llava : update clip.h (#7580)

overriden -> overridden

.github/ISSUE_TEMPLATE/01-bug-low.yml

@@ -0,0 +1,50 @@
+name: Low Severity Bugs
+description: Used to report low severity bugs in llama.cpp (e.g. cosmetic issues, non critical UI glitches)
+title: "Bug: "
+labels: ["bug-unconfirmed", "low severity"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for taking the time to fill out this bug report!
+        Please include information about your system, the steps to reproduce the bug,
+        and the version of llama.cpp that you are using.
+        If possible, please provide a minimal code example that reproduces the bug.
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What happened?
+      description: Also tell us, what did you expect to happen?
+      placeholder: Tell us what you see!
+    validations:
+      required: true
+  - type: textarea
+    id: version
+    attributes:
+      label: Name and Version
+      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
+      placeholder: |
+        $./main --version
+        version: 2999 (42b4109e)
+        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
+    validations:
+      required: true
+  - type: dropdown
+    id: operating-system
+    attributes:
+      label: What operating system are you seeing the problem on?
+      multiple: true
+      options:
+        - Linux
+        - Mac
+        - Windows
+        - BSD
+        - Other? (Please let us know in description)
+    validations:
+      required: false
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant log output
+      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
+      render: shell

.github/ISSUE_TEMPLATE/02-bug-medium.yml

@@ -0,0 +1,50 @@
+name: Medium Severity Bug
+description: Used to report medium severity bugs in llama.cpp (e.g. Malfunctioning Features but generally still useable)
+title: "Bug: "
+labels: ["bug-unconfirmed", "medium severity"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for taking the time to fill out this bug report!
+        Please include information about your system, the steps to reproduce the bug,
+        and the version of llama.cpp that you are using.
+        If possible, please provide a minimal code example that reproduces the bug.
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What happened?
+      description: Also tell us, what did you expect to happen?
+      placeholder: Tell us what you see!
+    validations:
+      required: true
+  - type: textarea
+    id: version
+    attributes:
+      label: Name and Version
+      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
+      placeholder: |
+        $./main --version
+        version: 2999 (42b4109e)
+        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
+    validations:
+      required: true
+  - type: dropdown
+    id: operating-system
+    attributes:
+      label: What operating system are you seeing the problem on?
+      multiple: true
+      options:
+        - Linux
+        - Mac
+        - Windows
+        - BSD
+        - Other? (Please let us know in description)
+    validations:
+      required: false
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant log output
+      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
+      render: shell

.github/ISSUE_TEMPLATE/03-bug-high.yml

@@ -0,0 +1,50 @@
+name: High Severity Bug
+description: Used to report high severity bugs in llama.cpp (e.g. Malfunctioning features hindering important common workflow)
+title: "Bug: "
+labels: ["bug-unconfirmed", "high severity"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for taking the time to fill out this bug report!
+        Please include information about your system, the steps to reproduce the bug,
+        and the version of llama.cpp that you are using.
+        If possible, please provide a minimal code example that reproduces the bug.
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What happened?
+      description: Also tell us, what did you expect to happen?
+      placeholder: Tell us what you see!
+    validations:
+      required: true
+  - type: textarea
+    id: version
+    attributes:
+      label: Name and Version
+      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
+      placeholder: |
+        $./main --version
+        version: 2999 (42b4109e)
+        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
+    validations:
+      required: true
+  - type: dropdown
+    id: operating-system
+    attributes:
+      label: What operating system are you seeing the problem on?
+      multiple: true
+      options:
+        - Linux
+        - Mac
+        - Windows
+        - BSD
+        - Other? (Please let us know in description)
+    validations:
+      required: false
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant log output
+      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
+      render: shell

.github/ISSUE_TEMPLATE/04-bug-critical.yml

@@ -0,0 +1,50 @@
+name: Critical Severity Bug
+description: Used to report critical severity bugs in llama.cpp (e.g. Crashing, Corrupted, Dataloss)
+title: "Bug: "
+labels: ["bug-unconfirmed", "critical severity"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for taking the time to fill out this bug report!
+        Please include information about your system, the steps to reproduce the bug,
+        and the version of llama.cpp that you are using.
+        If possible, please provide a minimal code example that reproduces the bug.
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What happened?
+      description: Also tell us, what did you expect to happen?
+      placeholder: Tell us what you see!
+    validations:
+      required: true
+  - type: textarea
+    id: version
+    attributes:
+      label: Name and Version
+      description: Which executable and which version of our software are you running? (use `--version` to get a version string)
+      placeholder: |
+        $./main --version
+        version: 2999 (42b4109e)
+        built with cc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 for x86_64-linux-gnu
+    validations:
+      required: true
+  - type: dropdown
+    id: operating-system
+    attributes:
+      label: What operating system are you seeing the problem on?
+      multiple: true
+      options:
+        - Linux
+        - Mac
+        - Windows
+        - BSD
+        - Other? (Please let us know in description)
+    validations:
+      required: false
+  - type: textarea
+    id: logs
+    attributes:
+      label: Relevant log output
+      description: Please copy and paste any relevant log output. This will be automatically formatted into code, so no need for backticks.
+      render: shell

.github/ISSUE_TEMPLATE/05-enhancement.yml

@@ -0,0 +1,51 @@
+name: Enhancement template
+description: Used to request enhancements for llama.cpp
+title: "Feature Request: "
+labels: ["enhancement"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        [Please post your idea first in Discussion if there is not yet a consensus for this enhancement request. This will help to keep this issue tracker focused on enhancements that the community has agreed needs to be implemented.](https://github.com/ggerganov/llama.cpp/discussions/categories/ideas)
+
+  - type: checkboxes
+    id: prerequisites
+    attributes:
+      label: Prerequisites
+      description: Please confirm the following before submitting your enhancement request.
+      options:
+        - label: I am running the latest code. Mention the version if possible as well.
+          required: true
+        - label: I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
+          required: true
+        - label: I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
+          required: true
+        - label: I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new and useful enhancement to share.
+          required: true
+
+  - type: textarea
+    id: feature-description
+    attributes:
+      label: Feature Description
+      description: Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement.
+      placeholder: Detailed description of the enhancement
+    validations:
+      required: true
+
+  - type: textarea
+    id: motivation
+    attributes:
+      label: Motivation
+      description: Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users.
+      placeholder: Explanation of why this feature is needed and its benefits
+    validations:
+      required: true
+
+  - type: textarea
+    id: possible-implementation
+    attributes:
+      label: Possible Implementation
+      description: If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better.
+      placeholder: Detailed description of potential implementation
+    validations:
+      required: false

.github/ISSUE_TEMPLATE/06-question.yml

@@ -0,0 +1,38 @@
+name: Question template
+description: Used to ask questions about llama.cpp
+title: "Question: "
+labels: ["question"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        [Please search your question first in Discussion if you got a common general question.](https://github.com/ggerganov/llama.cpp/discussions/categories/q-a)
+
+  - type: checkboxes
+    id: prerequisites
+    attributes:
+      label: Prerequisites
+      description: Please confirm the following before submitting your question.
+      options:
+        - label: I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
+          required: true
+        - label: I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new useful question to share that cannot be answered within Discussions.
+          required: true
+
+  - type: textarea
+    id: background-description
+    attributes:
+      label: Background Description
+      description: Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an question.
+      placeholder: Detailed description of your question
+    validations:
+      required: true
+
+  - type: textarea
+    id: possible-answer
+    attributes:
+      label: Possible Answer
+      description: If you have some idea of possible answers you want to confirm, that would also be appreciated.
+      placeholder: Your idea of possible answers
+    validations:
+      required: false

.github/ISSUE_TEMPLATE/bug.md

@@ -1,11 +0,0 @@
----
-name: Bug template
-about: Used to report bugs in llama.cpp
-labels: ["bug-unconfirmed"]
-assignees: ''
-
----
-
-Please include information about your system, the steps to reproduce the bug, and the version of llama.cpp that you are using. If possible, please provide a minimal code example that reproduces the bug.
-
-If the bug concerns the server, please try to reproduce it first using the [server test scenario framework](https://github.com/ggerganov/llama.cpp/tree/master/examples/server/tests).

.github/ISSUE_TEMPLATE/enhancement.md

@@ -1,28 +0,0 @@
----
-name: Enhancement template
-about: Used to request enhancements for llama.cpp
-labels: ["enhancement"]
-assignees: ''
-
----
-
-# Prerequisites
-
-Please answer the following questions for yourself before submitting an issue.
-
-- [ ] I am running the latest code. Development is very rapid so there are no tagged versions as of now.
-- [ ] I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
-- [ ] I [searched using keywords relevant to my issue](https://docs.github.com/en/issues/tracking-your-work-with-issues/filtering-and-searching-issues-and-pull-requests) to make sure that I am creating a new issue that is not already open (or closed).
-- [ ] I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new bug or useful enhancement to share.
-
-# Feature Description
-
-Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do as an enhancement.
-
-# Motivation
-
-Please provide a detailed written description of reasons why this feature is necessary and how it is useful to `llama.cpp` users.
-
-# Possible Implementation
-
-If you have an idea as to how it can be implemented, please write a detailed description. Feel free to give links to external sources or share visuals that might be helpful to understand the details better.

.github/labeler.yml

@@ -1,5 +1,16 @@
 # https://github.com/actions/labeler
-
+Kompute:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-kompute.h
+            - ggml-kompute.cpp
+            - README-kompute.md
+Apple Metal:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-metal.h
+            - ggml-metal.cpp
+            - README-metal.md
 SYCL:
     - changed-files:
         - any-glob-to-any-file:
@@ -9,6 +20,7 @@ SYCL:
 Nvidia GPU:
     - changed-files:
         - any-glob-to-any-file:
+            - ggml-cuda.h
             - ggml-cuda/**
 Vulkan:
     - changed-files:
@@ -62,6 +74,8 @@ server:
 ggml:
     - changed-files:
         - any-glob-to-any-file:
+            - ggml.c
+            - ggml.h
             - ggml-*.c
             - ggml-*.h
             - ggml-cuda/**
@@ -71,3 +85,6 @@ nix:
             - "**/*.nix"
             - .github/workflows/nix-*.yml
             - .devops/nix/nixpkgs-instances.nix
+embedding:
+    - changed-files:
+        - any-glob-to-any-file: examples/embedding/

.github/workflows/docker.yml

@@ -42,8 +42,9 @@ jobs:
           - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "server-rocm", dockerfile: ".devops/server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          - { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" }
-          - { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
+          # TODO: Disabled due to build issues https://github.com/ggerganov/llama.cpp/issues/7507
+          #- { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" }
+          #- { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v4

CMakeLists.txt

@@ -72,6 +72,7 @@ else()
     set(INS_ENB ON)
 endif()
 
+option(LLAMA_SVE                             "llama: enable SVE"                                OFF)
 option(LLAMA_AVX                             "llama: enable AVX"                                ${INS_ENB})
 option(LLAMA_AVX2                            "llama: enable AVX2"                               ${INS_ENB})
 option(LLAMA_AVX512                          "llama: enable AVX512"                             OFF)
@@ -124,7 +125,6 @@ set(LLAMA_METAL_MACOSX_VERSION_MIN "" CACHE STRING
 set(LLAMA_METAL_STD "" CACHE STRING          "llama: metal standard version (-std flag)")
 option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
 option(LLAMA_RPC                             "llama: use RPC"                                   OFF)
-option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
 set(LLAMA_SYCL_TARGET   "INTEL" CACHE STRING "llama: sycl target device")
@@ -384,10 +384,6 @@ if (LLAMA_LLAMAFILE)
     set(GGML_SOURCES_LLAMAFILE sgemm.cpp)
 endif()
 
-if (LLAMA_QKK_64)
-    add_compile_definitions(GGML_QKK_64)
-endif()
-
 if (LLAMA_CUBLAS)
     message(WARNING "LLAMA_CUBLAS is deprecated and will be removed in the future.\nUse LLAMA_CUDA instead")
     set(LLAMA_CUDA ON)
@@ -1045,6 +1041,9 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
             # Raspberry Pi 3, 4, Zero 2 (32-bit)
             list(APPEND ARCH_FLAGS -mno-unaligned-access)
         endif()
+        if (LLAMA_SVE)
+            list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
+        endif()
     endif()
 elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
         (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND

CMakePresets.json

@@ -1,4 +1,4 @@
-{
+{
   "version": 4,
   "configurePresets": [
     {
@@ -40,6 +40,10 @@
 
     { "name": "arm64-windows-msvc-debug"  , "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
     { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "release" ] },
-    { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "release", "static" ] }
+    { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "release", "static" ] },
+
+    { "name": "x64-windows-msvc-debug"  , "inherits": [ "base", "debug"   ] },
+    { "name": "x64-windows-msvc-release", "inherits": [ "base", "release" ] },
+    { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "release", "static" ] }
   ]
 }

Makefile

@@ -389,10 +389,6 @@ else
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif
 
-ifdef LLAMA_QKK_64
-	MK_CPPFLAGS += -DGGML_QKK_64
-endif
-
 ifndef LLAMA_NO_ACCELERATE
 	# Mac OS - include Accelerate framework.
 	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
@@ -445,6 +441,9 @@ endif # JETSON_EOL_MODULE_DETECT
 ifdef LLAMA_DEBUG
 	MK_NVCCFLAGS += -lineinfo
 endif # LLAMA_DEBUG
+ifdef LLAMA_CUDA_DEBUG
+	MK_NVCCFLAGS += --device-debug
+endif # LLAMA_CUDA_DEBUG
 ifdef LLAMA_CUDA_NVCC
 	NVCC = $(CCACHE) $(LLAMA_CUDA_NVCC)
 else

README.md

@@ -127,6 +127,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
+- [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md))
 
@@ -140,6 +141,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
 - [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
 - [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
+- [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
 
 **HTTP server**
 
@@ -201,6 +203,10 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
+**Tools:**
+
+- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
+
 ---
 
 Here is a typical run using LLaMA v2 13B on M2 Ultra:

ci/run.sh

@@ -202,12 +202,15 @@ function gg_sum_test_scripts_release {
 }
 
 function gg_get_model {
-    local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf"
-    local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
-    if [[ -s $gguf_3b ]]; then
-        echo -n "$gguf_3b"
-    elif [[ -s $gguf_7b ]]; then
-        echo -n "$gguf_7b"
+    local gguf_0="$MNT/models/pythia/1.4B/ggml-model-f16.gguf"
+    local gguf_1="$MNT/models/pythia/2.8B/ggml-model-f16.gguf"
+    local gguf_2="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
+    if [[ -s $gguf_0 ]]; then
+        echo -n "$gguf_0"
+    elif [[ -s $gguf_1 ]]; then
+        echo -n "$gguf_1"
+    elif [[ -s $gguf_2 ]]; then
+        echo -n "$gguf_2"
     else
         echo >&2 "No model found. Can't run gg_run_ctest_with_model."
         exit 1
@@ -256,139 +259,6 @@ function gg_sum_ctest_with_model_release {
     gg_printf '```\n'
 }
 
-# open_llama_3b_v2
-
-function gg_run_open_llama_3b_v2 {
-    cd ${SRC}
-
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/config.json
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/tokenizer.model
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/tokenizer_config.json
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/special_tokens_map.json
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/pytorch_model.bin
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/generation_config.json
-
-    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
-    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
-    head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
-
-    path_models="../models-mnt/open-llama/3B-v2"
-    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
-
-    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
-
-    set -e
-
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_QKK_64=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                                             ) 2>&1 | tee -a $OUT/${ci}-make.log
-
-    python3 ../convert.py ${path_models}
-
-    model_f16="${path_models}/ggml-model-f16.gguf"
-    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
-    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
-    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
-    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
-    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
-    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
-    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
-    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
-    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
-    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
-
-    wiki_test_60="${path_wiki}/wiki.test-60.raw"
-
-    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
-    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
-    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
-    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
-    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
-    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
-    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
-    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
-    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
-    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
-
-    (time ./bin/main --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/main --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/main --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/main --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/main --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/main --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/main --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/main --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/main --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
-
-    (time ./bin/save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-
-    function check_ppl {
-        qnt="$1"
-        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
-            return 20
-        fi
-
-        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
-        return 0
-    }
-
-    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-
-    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
-
-    set +e
-}
-
-function gg_sum_open_llama_3b_v2 {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'OpenLLaMA 3B-v2:\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
-    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
-    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
-    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
-    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
-    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
-    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
-    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
-    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
-    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
-    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
-    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
-    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
-}
-
 # open_llama_7b_v2
 # requires: GG_BUILD_CUDA
 
@@ -417,7 +287,7 @@ function gg_run_open_llama_7b_v2 {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../convert.py ${path_models}
+    python3 ../convert.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -526,6 +396,272 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
 }
 
+# pythia_1.4b
+
+function gg_run_pythia_1_4b {
+    cd ${SRC}
+
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/config.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer_config.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/special_tokens_map.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/resolve/main/pytorch_model.bin
+
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+    head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
+
+    path_models="../models-mnt/pythia/1.4B"
+    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
+    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
+    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
+    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
+    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
+    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
+    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
+    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
+    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
+
+    wiki_test_60="${path_wiki}/wiki.test-60.raw"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/main --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
+    (time ./bin/save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+
+    function check_ppl {
+        qnt="$1"
+        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
+        return 0
+    }
+
+    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+   #check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
+    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+
+    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
+
+    set +e
+}
+
+function gg_sum_pythia_1_4b {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Pythia 1.4B:\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
+    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
+    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
+    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
+    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
+    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
+    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
+    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
+    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
+}
+
+# pythia_2_8b
+# requires: GG_BUILD_CUDA
+
+function gg_run_pythia_2_8b {
+    cd ${SRC}
+
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/config.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer_config.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/special_tokens_map.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/resolve/main/pytorch_model.bin
+
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+
+    path_models="../models-mnt/pythia/2.8B"
+    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
+    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
+    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
+    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
+    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
+    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
+    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
+    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
+    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
+
+    wiki_test="${path_wiki}/wiki.test.raw"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/main --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
+    (time ./bin/save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+
+    function check_ppl {
+        qnt="$1"
+        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
+        return 0
+    }
+
+    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+   #check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
+    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+
+    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
+
+    set +e
+}
+
+function gg_sum_pythia_2_8b {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Pythia 2.8B:\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
+    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
+    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
+    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
+    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
+    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
+    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
+    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
+    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
+}
+
 # bge-small
 
 function gg_run_embd_bge_small {
@@ -552,7 +688,7 @@ function gg_run_embd_bge_small {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../convert-hf-to-gguf.py ${path_models}
+    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -606,9 +742,10 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
 
     if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
         if [ -z ${GG_BUILD_CUDA} ]; then
-            test $ret -eq 0 && gg_run open_llama_3b_v2
+            test $ret -eq 0 && gg_run pythia_1_4b
         else
-            test $ret -eq 0 && gg_run open_llama_7b_v2
+            test $ret -eq 0 && gg_run pythia_2_8b
+            #test $ret -eq 0 && gg_run open_llama_7b_v2
         fi
         test $ret -eq 0 && gg_run ctest_with_model_debug
         test $ret -eq 0 && gg_run ctest_with_model_release

common/common.cpp

@@ -904,6 +904,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.interactive_specials = true;
         return true;
     }
+    if (arg == "--special") {
+        params.special = true;
+        return true;
+    }
     if (arg == "--embedding") {
         params.embedding = true;
         return true;
@@ -1362,6 +1366,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
     printf("  -h, --help            show this help message and exit\n");
     printf("  --version             show version and build info\n");
     printf("  -i, --interactive     run in interactive mode\n");
+    printf("  --special             special tokens output enabled\n");
     printf("  --interactive-specials allow special tokens in user text, in interactive mode\n");
     printf("  --interactive-first   run in interactive mode and wait for input right away\n");
     printf("  -cnv, --conversation  run in conversation mode (does not print special tokens and suffix/prefix)\n");
@@ -1855,11 +1860,15 @@ bool fs_create_directory_with_parents(const std::string & path) {
 
 std::string fs_get_cache_directory() {
     std::string cache_directory = "";
+    auto ensure_trailing_slash = [](std::string p) {
+        // Make sure to add trailing slash
+        if (p.back() != DIRECTORY_SEPARATOR) {
+            p += DIRECTORY_SEPARATOR;
+        }
+        return p;
+    };
     if (getenv("LLAMA_CACHE")) {
         cache_directory = std::getenv("LLAMA_CACHE");
-        if (cache_directory.back() != DIRECTORY_SEPARATOR) {
-            cache_directory += DIRECTORY_SEPARATOR;
-        }
     } else {
 #ifdef __linux__
         if (std::getenv("XDG_CACHE_HOME")) {
@@ -1870,12 +1879,12 @@ std::string fs_get_cache_directory() {
 #elif defined(__APPLE__)
         cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
-        cache_directory = std::getenv("APPDATA");
+        cache_directory = std::getenv("LOCALAPPDATA");
 #endif // __linux__
+        cache_directory = ensure_trailing_slash(cache_directory);
         cache_directory += "llama.cpp";
-        cache_directory += DIRECTORY_SEPARATOR;
     }
-    return cache_directory;
+    return ensure_trailing_slash(cache_directory);
 }
 
 
@@ -2840,6 +2849,7 @@ void yaml_dump_non_result_info(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "cpu_has_fma: %s\n",         ggml_cpu_has_fma()         ? "true" : "false");
     fprintf(stream, "cpu_has_gpublas: %s\n",     ggml_cpu_has_gpublas()     ? "true" : "false");
     fprintf(stream, "cpu_has_neon: %s\n",        ggml_cpu_has_neon()        ? "true" : "false");
+    fprintf(stream, "cpu_has_sve: %s\n",         ggml_cpu_has_sve()         ? "true" : "false");
     fprintf(stream, "cpu_has_f16c: %s\n",        ggml_cpu_has_f16c()        ? "true" : "false");
     fprintf(stream, "cpu_has_fp16_va: %s\n",     ggml_cpu_has_fp16_va()     ? "true" : "false");
     fprintf(stream, "cpu_has_wasm_simd: %s\n",   ggml_cpu_has_wasm_simd()   ? "true" : "false");

common/common.h

@@ -146,6 +146,7 @@ struct gpt_params {
     bool use_color         = false; // use color to distinguish generations and inputs
     bool interactive       = false; // interactive mode
     bool interactive_specials = false; // whether to allow special tokens from user, during interactive mode
+    bool special           = false; // enable special token output
     bool conversation      = false; // conversation mode (does not print special tokens and suffix/prefix)
     bool chatml            = false; // chatml mode (used for models trained on chatml syntax)
     bool prompt_cache_all  = false; // save user input and generations to prompt cache

common/train.cpp

@@ -1052,7 +1052,7 @@ struct train_params_common get_default_train_params_common() {
 
     params.custom_n_ctx = false;
 
-    params.use_flash              = true;
+    params.use_flash              = false;
     params.use_checkpointing      = true;
 
     params.sample_start           = "";

convert-hf-to-gguf-update.py

@@ -81,6 +81,7 @@ models = [
     {"name": "jina-v2-en",     "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
     {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
     {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
+    {"name": "smaug-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
 ]
 
 

convert-hf-to-gguf.py

@@ -313,11 +313,10 @@ class Model:
                         data = data.astype(np.float32)
                     data_qtype = gguf.GGMLQuantizationType.F32
 
-                block_size, type_size = gguf.GGML_QUANT_SIZES[data_qtype]
+                shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape
+
                 # reverse shape to make it similar to the internal ggml dimension order
-                shape_str = f"""{{{', '.join(str(n) for n in reversed(
-                    (*data.shape[:-1], data.shape[-1] * data.dtype.itemsize // type_size * block_size))
-                )}}}"""
+                shape_str = f"{{{', '.join(str(n) for n in reversed(shape))}}}"
 
                 # n_dims is implicit in the shape
                 logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
@@ -474,6 +473,9 @@ class Model:
         if chkhsh == "27949a2493fc4a9f53f5b9b029c82689cfbe5d3a1929bb25e043089e28466de6":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-de
             res = "jina-v2-de"
+        if chkhsh == "c136ed14d01c2745d4f60a9596ae66800e2b61fa45643e72436041855ad4089d":
+            # ref: https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct
+            res = "smaug-bpe"
 
         if res is None:
             logger.warning("\n")
@@ -673,6 +675,44 @@ class GPTNeoXModel(Model):
         self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
+        n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
+
+        tensors: list[tuple[str, Tensor]] = []
+
+        if re.match(r"gpt_neox\.layers\.\d+\.attention\.query_key_value\.weight", name):
+            # Map bloom-style qkv_linear to gpt-style qkv_linear
+            # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
+            # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312  # noqa
+            qkv_weights = data_torch.reshape((n_head, 3, n_embed // n_head, n_embed))
+            data_torch = torch.cat(
+                (
+                    qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
+                    qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
+                    qkv_weights[:, 2, :, :].reshape((-1, n_embed)),
+                ),
+                dim=0,
+            )
+            logger.info("re-format attention.linear_qkv.weight")
+        elif re.match(r"gpt_neox\.layers\.\d+\.attention\.query_key_value\.bias", name):
+            qkv_bias = data_torch.reshape((n_head, 3, n_embed // n_head))
+            data_torch = torch.cat(
+                (
+                    qkv_bias[:, 0, :].reshape((n_embed,)),
+                    qkv_bias[:, 1, :].reshape((n_embed,)),
+                    qkv_bias[:, 2, :].reshape((n_embed,)),
+                ),
+                dim=0,
+            )
+            logger.info("re-format attention.linear_qkv.bias")
+
+        tensors.append((self.map_tensor_name(name), data_torch))
+
+        return tensors
+
 
 @Model.register("BloomForCausalLM")
 class BloomModel(Model):
@@ -2355,7 +2395,8 @@ class CommandR2Model(Model):
 
         # max_position_embeddings = 8192 in config.json but model was actually
         # trained on 128k context length
-        self.hparams["max_position_embeddings"] = self.hparams["model_max_length"]
+        # aya-23 models don't have model_max_length specified
+        self.hparams["max_position_embeddings"] = self.find_hparam(["model_max_length", "max_position_embeddings"])
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -2428,6 +2469,157 @@ class JinaBertV2Model(BertModel):
         self.gguf_writer.add_add_eos_token(True)
 
 
+@Model.register("ArcticForCausalLM")
+class ArcticModel(Model):
+    model_arch = gguf.MODEL_ARCH.ARCTIC
+
+    def set_vocab(self):
+        # The reason for using a custom implementation here is that the
+        # snowflake-arctic-instruct model redefined tokens 31998 and 31999 from
+        # tokenizer.model and used them as BOS and EOS instead of adding new tokens.
+        from sentencepiece import SentencePieceProcessor
+
+        tokenizer_path = self.dir_model / 'tokenizer.model'
+
+        if not tokenizer_path.is_file():
+            logger.error(f'Error: Missing {tokenizer_path}')
+            sys.exit(1)
+
+        # Read the whole vocabulary from the tokenizer.model file
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        # Use the added_tokens_decoder field from tokeniser_config.json as the source
+        # of information about added/redefined tokens and modify them accordingly.
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+
+                if "added_tokens_decoder" in tokenizer_config_json:
+                    added_tokens_decoder = tokenizer_config_json["added_tokens_decoder"]
+                    for token_id, token_json in added_tokens_decoder.items():
+                        token_id = int(token_id)
+                        if (token_id >= vocab_size):
+                            logger.debug(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                            continue
+
+                        token_content = token_json["content"]
+                        token_type = SentencePieceTokenTypes.USER_DEFINED
+                        token_score = -10000.0
+
+                        # Map unk_token to UNKNOWN, other special tokens to CONTROL
+                        # Set the score to 0.0 as in the original tokenizer.model
+                        if ("special" in token_json) and token_json["special"]:
+                            if token_content == tokenizer_config_json["unk_token"]:
+                                token_type = SentencePieceTokenTypes.UNKNOWN
+                            else:
+                                token_type = SentencePieceTokenTypes.CONTROL
+                            token_score = 0.0
+
+                        logger.info(f"Setting added token {token_id} to '{token_content}' (type: {token_type}, score: {token_score:.2f})")
+                        tokens[token_id] = token_content.encode("utf-8")
+                        toktypes[token_id] = token_type
+                        scores[token_id] = token_score
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+
+        if name.endswith("q_proj.weight"):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+        if name.endswith("k_proj.weight"):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+
+        # process the experts separately
+        if name.find("block_sparse_moe.experts") != -1:
+            n_experts = self.hparams["num_local_experts"]
+
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for wid in ["w1", "w2", "w3"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"layers.{bid}.feed_forward.experts.{wid}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def write_tensors(self):
+        super().write_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 ###### CONVERSION LOGIC ######
 
 

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -774,7 +774,7 @@ static struct train_params get_default_train_params() {
 
     params.samples_start_after_nl = false;
     params.use_adam               = true;
-    params.use_flash              = true;
+    params.use_flash              = false;
     params.use_scratch            = true;
 
     // only adam

examples/finetune/finetune.cpp

@@ -643,7 +643,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         struct ggml_tensor * t15 = ggml_permute      (ctx, t12, 0, 3, 1, 2);                         set_name(t15, "t15");     assert_shape_4d(t15, N, n_embd_head, n_head_kv, n_batch);
         struct ggml_tensor * t16;
         if (enable_flash_attn) {
-            t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                         set_name(t16, "t16");     assert_shape_4d(t16, n_embd_head, N, n_head, n_batch);
+            GGML_ASSERT(false && "TODO: ggml_flash_attn_ext() not yet supported");
+            //t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                         set_name(t16, "t16");     assert_shape_4d(t16, n_embd_head, N, n_head, n_batch);
         } else {
             struct ggml_tensor * t16_0 = ggml_mul_mat              (ctx, t14, t13);                  set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
             struct ggml_tensor * t16_1 = ggml_scale_inplace        (ctx, t16_0, kv_scale);           set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);

examples/llama.android/app/build.gradle.kts

@@ -7,8 +7,6 @@ android {
     namespace = "com.example.llama"
     compileSdk = 34
 
-    ndkVersion = "26.1.10909125"
-
     defaultConfig {
         applicationId = "com.example.llama"
         minSdk = 33
@@ -20,17 +18,6 @@ android {
         vectorDrawables {
             useSupportLibrary = true
         }
-        ndk {
-            // Add NDK properties if wanted, e.g.
-            // abiFilters += listOf("arm64-v8a")
-        }
-        externalNativeBuild {
-            cmake {
-                arguments += "-DCMAKE_BUILD_TYPE=Release"
-                cppFlags += listOf()
-                arguments += listOf()
-            }
-        }
     }
 
     buildTypes {
@@ -55,17 +42,6 @@ android {
     composeOptions {
         kotlinCompilerExtensionVersion = "1.5.1"
     }
-    packaging {
-        resources {
-            excludes += "/META-INF/{AL2.0,LGPL2.1}"
-        }
-    }
-    externalNativeBuild {
-        cmake {
-            path = file("src/main/cpp/CMakeLists.txt")
-            version = "3.22.1"
-        }
-    }
 }
 
 dependencies {
@@ -78,6 +54,7 @@ dependencies {
     implementation("androidx.compose.ui:ui-graphics")
     implementation("androidx.compose.ui:ui-tooling-preview")
     implementation("androidx.compose.material3:material3")
+    implementation(project(":llama"))
     testImplementation("junit:junit:4.13.2")
     androidTestImplementation("androidx.test.ext:junit:1.1.5")
     androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1")

examples/llama.android/app/src/main/java/com/example/llama/MainViewModel.kt

@@ -1,5 +1,6 @@
 package com.example.llama
 
+import android.llama.cpp.LLamaAndroid
 import android.util.Log
 import androidx.compose.runtime.getValue
 import androidx.compose.runtime.mutableStateOf
@@ -9,7 +10,7 @@ import androidx.lifecycle.viewModelScope
 import kotlinx.coroutines.flow.catch
 import kotlinx.coroutines.launch
 
-class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
+class MainViewModel(private val llamaAndroid: LLamaAndroid = LLamaAndroid.instance()): ViewModel() {
     companion object {
         @JvmStatic
         private val NanosPerSecond = 1_000_000_000.0
@@ -28,7 +29,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
 
         viewModelScope.launch {
             try {
-                llm.unload()
+                llamaAndroid.unload()
             } catch (exc: IllegalStateException) {
                 messages += exc.message!!
             }
@@ -44,7 +45,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
         messages += ""
 
         viewModelScope.launch {
-            llm.send(text)
+            llamaAndroid.send(text)
                 .catch {
                     Log.e(tag, "send() failed", it)
                     messages += it.message!!
@@ -57,7 +58,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
         viewModelScope.launch {
             try {
                 val start = System.nanoTime()
-                val warmupResult = llm.bench(pp, tg, pl, nr)
+                val warmupResult = llamaAndroid.bench(pp, tg, pl, nr)
                 val end = System.nanoTime()
 
                 messages += warmupResult
@@ -70,7 +71,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
                     return@launch
                 }
 
-                messages += llm.bench(512, 128, 1, 3)
+                messages += llamaAndroid.bench(512, 128, 1, 3)
             } catch (exc: IllegalStateException) {
                 Log.e(tag, "bench() failed", exc)
                 messages += exc.message!!
@@ -81,7 +82,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
     fun load(pathToModel: String) {
         viewModelScope.launch {
             try {
-                llm.load(pathToModel)
+                llamaAndroid.load(pathToModel)
                 messages += "Loaded $pathToModel"
             } catch (exc: IllegalStateException) {
                 Log.e(tag, "load() failed", exc)

examples/llama.android/build.gradle.kts

@@ -2,4 +2,5 @@
 plugins {
     id("com.android.application") version "8.2.0" apply false
     id("org.jetbrains.kotlin.android") version "1.9.0" apply false
+    id("com.android.library") version "8.2.0" apply false
 }

examples/llama.android/llama/.gitignore

@@ -0,0 +1 @@
+/build

examples/llama.android/llama/CMakeLists.txt

@@ -42,7 +42,7 @@ add_subdirectory(../../../../../../ build-llama)
 # used in the AndroidManifest.xml file.
 add_library(${CMAKE_PROJECT_NAME} SHARED
     # List C/C++ source files with relative paths to this CMakeLists.txt.
-    llama-android.cpp)
+        llama-android.cpp)
 
 # Specifies libraries CMake should link to your target library. You
 # can link libraries from various origins, such as libraries defined in this

examples/llama.android/llama/build.gradle.kts

@@ -0,0 +1,68 @@
+plugins {
+    id("com.android.library")
+    id("org.jetbrains.kotlin.android")
+}
+
+android {
+    namespace = "android.llama.cpp"
+    compileSdk = 34
+
+    defaultConfig {
+        minSdk = 33
+
+        testInstrumentationRunner = "androidx.test.runner.AndroidJUnitRunner"
+        consumerProguardFiles("consumer-rules.pro")
+        ndk {
+            // Add NDK properties if wanted, e.g.
+            // abiFilters += listOf("arm64-v8a")
+        }
+        externalNativeBuild {
+            cmake {
+                arguments += "-DCMAKE_BUILD_TYPE=Release"
+                cppFlags += listOf()
+                arguments += listOf()
+
+                cppFlags("")
+            }
+        }
+    }
+
+    buildTypes {
+        release {
+            isMinifyEnabled = false
+            proguardFiles(
+                getDefaultProguardFile("proguard-android-optimize.txt"),
+                "proguard-rules.pro"
+            )
+        }
+    }
+    externalNativeBuild {
+        cmake {
+            path("src/main/cpp/CMakeLists.txt")
+            version = "3.22.1"
+        }
+    }
+    compileOptions {
+        sourceCompatibility = JavaVersion.VERSION_1_8
+        targetCompatibility = JavaVersion.VERSION_1_8
+    }
+    kotlinOptions {
+        jvmTarget = "1.8"
+    }
+
+    packaging {
+        resources {
+            excludes += "/META-INF/{AL2.0,LGPL2.1}"
+        }
+    }
+}
+
+dependencies {
+
+    implementation("androidx.core:core-ktx:1.12.0")
+    implementation("androidx.appcompat:appcompat:1.6.1")
+    implementation("com.google.android.material:material:1.11.0")
+    testImplementation("junit:junit:4.13.2")
+    androidTestImplementation("androidx.test.ext:junit:1.1.5")
+    androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1")
+}

examples/llama.android/llama/proguard-rules.pro

@@ -0,0 +1,21 @@
+# Add project specific ProGuard rules here.
+# You can control the set of applied configuration files using the
+# proguardFiles setting in build.gradle.
+#
+# For more details, see
+#   http://developer.android.com/guide/developing/tools/proguard.html
+
+# If your project uses WebView with JS, uncomment the following
+# and specify the fully qualified class name to the JavaScript interface
+# class:
+#-keepclassmembers class fqcn.of.javascript.interface.for.webview {
+#   public *;
+#}
+
+# Uncomment this to preserve the line number information for
+# debugging stack traces.
+#-keepattributes SourceFile,LineNumberTable
+
+# If you keep the line number information, uncomment this to
+# hide the original source file name.
+#-renamesourcefileattribute SourceFile

examples/llama.android/llama/src/androidTest/java/android/llama/cpp/ExampleInstrumentedTest.kt

@@ -0,0 +1,24 @@
+package android.llama.cpp
+
+import androidx.test.platform.app.InstrumentationRegistry
+import androidx.test.ext.junit.runners.AndroidJUnit4
+
+import org.junit.Test
+import org.junit.runner.RunWith
+
+import org.junit.Assert.*
+
+/**
+ * Instrumented test, which will execute on an Android device.
+ *
+ * See [testing documentation](http://d.android.com/tools/testing).
+ */
+@RunWith(AndroidJUnit4::class)
+class ExampleInstrumentedTest {
+    @Test
+    fun useAppContext() {
+        // Context of the app under test.
+        val appContext = InstrumentationRegistry.getInstrumentation().targetContext
+        assertEquals("android.llama.cpp.test", appContext.packageName)
+    }
+}

examples/llama.android/llama/src/main/AndroidManifest.xml

@@ -0,0 +1,4 @@
+<?xml version="1.0" encoding="utf-8"?>
+<manifest xmlns:android="http://schemas.android.com/apk/res/android">
+
+</manifest>

examples/llama.android/llama/src/main/cpp/CMakeLists.txt

@@ -0,0 +1,49 @@
+# For more information about using CMake with Android Studio, read the
+# documentation: https://d.android.com/studio/projects/add-native-code.html.
+# For more examples on how to use CMake, see https://github.com/android/ndk-samples.
+
+# Sets the minimum CMake version required for this project.
+cmake_minimum_required(VERSION 3.22.1)
+
+# Declares the project name. The project name can be accessed via ${ PROJECT_NAME},
+# Since this is the top level CMakeLists.txt, the project name is also accessible
+# with ${CMAKE_PROJECT_NAME} (both CMake variables are in-sync within the top level
+# build script scope).
+project("llama-android")
+
+include(FetchContent)
+FetchContent_Declare(
+        llama
+        GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
+        GIT_TAG        master
+)
+
+# Also provides "common"
+FetchContent_MakeAvailable(llama)
+
+# Creates and names a library, sets it as either STATIC
+# or SHARED, and provides the relative paths to its source code.
+# You can define multiple libraries, and CMake builds them for you.
+# Gradle automatically packages shared libraries with your APK.
+#
+# In this top level CMakeLists.txt, ${CMAKE_PROJECT_NAME} is used to define
+# the target library name; in the sub-module's CMakeLists.txt, ${PROJECT_NAME}
+# is preferred for the same purpose.
+#
+# In order to load a library into your app from Java/Kotlin, you must call
+# System.loadLibrary() and pass the name of the library defined here;
+# for GameActivity/NativeActivity derived applications, the same library name must be
+# used in the AndroidManifest.xml file.
+add_library(${CMAKE_PROJECT_NAME} SHARED
+        # List C/C++ source files with relative paths to this CMakeLists.txt.
+        llama-android.cpp)
+
+# Specifies libraries CMake should link to your target library. You
+# can link libraries from various origins, such as libraries defined in this
+# build script, prebuilt third-party libraries, or Android system libraries.
+target_link_libraries(${CMAKE_PROJECT_NAME}
+        # List libraries link to the target library
+        llama
+        common
+        android
+        log)

examples/llama.android/llama/src/main/cpp/llama-android.cpp

@@ -81,7 +81,7 @@ static void log_callback(ggml_log_level level, const char * fmt, void * data) {
 
 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) {
+Java_android_llama_cpp_LLamaAndroid_load_1model(JNIEnv *env, jobject, jstring filename) {
     llama_model_params model_params = llama_model_default_params();
 
     auto path_to_model = env->GetStringUTFChars(filename, 0);
@@ -101,13 +101,13 @@ Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) {
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1model(JNIEnv *, jobject, jlong model) {
+Java_android_llama_cpp_LLamaAndroid_free_1model(JNIEnv *, jobject, jlong model) {
     llama_free_model(reinterpret_cast<llama_model *>(model));
 }
 
 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) {
+Java_android_llama_cpp_LLamaAndroid_new_1context(JNIEnv *env, jobject, jlong jmodel) {
     auto model = reinterpret_cast<llama_model *>(jmodel);
 
     if (!model) {
@@ -139,25 +139,25 @@ Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) {
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1context(JNIEnv *, jobject, jlong context) {
+Java_android_llama_cpp_LLamaAndroid_free_1context(JNIEnv *, jobject, jlong context) {
     llama_free(reinterpret_cast<llama_context *>(context));
 }
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1free(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_backend_1free(JNIEnv *, jobject) {
     llama_backend_free();
 }
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_log_1to_1android(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_log_1to_1android(JNIEnv *, jobject) {
     llama_log_set(log_callback, NULL);
 }
 
 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_bench_1model(
+Java_android_llama_cpp_LLamaAndroid_bench_1model(
         JNIEnv *env,
         jobject,
         jlong context_pointer,
@@ -271,13 +271,13 @@ Java_com_example_llama_Llm_bench_1model(
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
+Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
     llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
 }
 
 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {
+Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {
 
     // Source: Copy of llama.cpp:llama_batch_init but heap-allocated.
 
@@ -313,19 +313,19 @@ Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint emb
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_backend_1init(JNIEnv *, jobject) {
     llama_backend_init();
 }
 
 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_system_1info(JNIEnv *env, jobject) {
+Java_android_llama_cpp_LLamaAndroid_system_1info(JNIEnv *env, jobject) {
     return env->NewStringUTF(llama_print_system_info());
 }
 
 extern "C"
 JNIEXPORT jint JNICALL
-Java_com_example_llama_Llm_completion_1init(
+Java_android_llama_cpp_LLamaAndroid_completion_1init(
         JNIEnv *env,
         jobject,
         jlong context_pointer,
@@ -376,7 +376,7 @@ Java_com_example_llama_Llm_completion_1init(
 
 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_completion_1loop(
+Java_android_llama_cpp_LLamaAndroid_completion_1loop(
         JNIEnv * env,
         jobject,
         jlong context_pointer,
@@ -438,6 +438,6 @@ Java_com_example_llama_Llm_completion_1loop(
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
+Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
     llama_kv_cache_clear(reinterpret_cast<llama_context *>(context));
 }

examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt

@@ -1,4 +1,4 @@
-package com.example.llama
+package android.llama.cpp
 
 import android.util.Log
 import kotlinx.coroutines.CoroutineDispatcher
@@ -10,7 +10,7 @@ import kotlinx.coroutines.withContext
 import java.util.concurrent.Executors
 import kotlin.concurrent.thread
 
-class Llm {
+class LLamaAndroid {
     private val tag: String? = this::class.simpleName
 
     private val threadLocalState: ThreadLocal<State> = ThreadLocal.withInitial { State.Idle }
@@ -165,8 +165,8 @@ class Llm {
         }
 
         // Enforce only one instance of Llm.
-        private val _instance: Llm = Llm()
+        private val _instance: LLamaAndroid = LLamaAndroid()
 
-        fun instance(): Llm = _instance
+        fun instance(): LLamaAndroid = _instance
     }
 }

examples/llama.android/llama/src/test/java/android/llama/cpp/ExampleUnitTest.kt

@@ -0,0 +1,17 @@
+package android.llama.cpp
+
+import org.junit.Test
+
+import org.junit.Assert.*
+
+/**
+ * Example local unit test, which will execute on the development machine (host).
+ *
+ * See [testing documentation](http://d.android.com/tools/testing).
+ */
+class ExampleUnitTest {
+    @Test
+    fun addition_isCorrect() {
+        assertEquals(4, 2 + 2)
+    }
+}

examples/llama.android/settings.gradle.kts

@@ -15,3 +15,4 @@ dependencyResolutionManagement {
 
 rootProject.name = "LlamaAndroid"
 include(":app")
+include(":llama")

examples/llava/clip.h

@@ -68,7 +68,7 @@ CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8
 /** interpret bytes as an image file with length bytes_length, and use the result to populate img */
 CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
 
-/** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */
+/** preprocess img and store the result in res_imgs, pad_to_square may be overridden to false depending on model configuration */
 CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
 
 CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);

examples/main/main.cpp

@@ -740,18 +740,26 @@ int main(int argc, char ** argv) {
         // display text
         if (input_echo && display) {
             for (auto id : embd) {
-                const std::string token_str = llama_token_to_piece(ctx, id, !params.conversation);
-                printf("%s", token_str.c_str());
+                const std::string token_str = llama_token_to_piece(ctx, id, params.special);
 
+                // Console/Stream Output
+                fprintf(stdout, "%s", token_str.c_str());
+
+                // Record Displayed Tokens To Log
+                // Note: Generated tokens are created one by one hence this check
                 if (embd.size() > 1) {
+                    // Incoming Requested Tokens
                     input_tokens.push_back(id);
                 } else {
+                    // Outgoing Generated Tokens
                     output_tokens.push_back(id);
                     output_ss << token_str;
                 }
+
+                fflush(stdout);
             }
-            fflush(stdout);
         }
+
         // reset color to default if there is no pending user input
         if (input_echo && (int) embd_inp.size() == n_consumed) {
             console::set_display(console::reset);

examples/server/public_simplechat/index.html

@@ -1,7 +1,7 @@
 <!DOCTYPE html>
 <html lang="en">
     <head>
-        <title>SimpleChat (LlamaCPP, ...) </title>
+        <title>SimpleChat LlamaCppEtal </title>
         <meta charset="UTF-8" />
         <meta name="viewport" content="width=device-width, initial-scale=1" />
         <meta name="message" content="Save Nature Save Earth" />
@@ -30,20 +30,17 @@
             <hr>
             <div class="sameline">
                 <label for="system-in">System</label>
-                <input type="text" name="system" id="system-in" class="flex-grow"/>
+                <input type="text" name="system" id="system-in" placeholder="e.g. you are a helpful ai assistant, who provides concise answers" class="flex-grow"/>
             </div>
 
             <hr>
             <div id="chat-div">
-                <p> Enter the system prompt above, before entering/submitting any user query.</p>
-                <p> Enter your text to the ai assistant below.</p>
-                <p> Use shift+enter for inserting enter.</p>
-                <p> Refresh the page to start over fresh.</p>
+                <p> You need to have javascript enabled.</p>
             </div>
 
             <hr>
             <div class="sameline">
-                <textarea id="user-in" class="flex-grow" rows="3"></textarea>
+                <textarea id="user-in" class="flex-grow" rows="3" placeholder="enter your query to the ai model here" ></textarea>
                 <button id="user-btn">submit</button>
             </div>
 

examples/server/public_simplechat/readme.md

@@ -14,11 +14,15 @@ own system prompts.
 The UI follows a responsive web design so that the layout can adapt to available display space in a usable
 enough manner, in general.
 
-NOTE: Given that the idea is for basic minimal testing, it doesnt bother with any model context length and
-culling of old messages from the chat.
+Allows developer/end-user to control some of the behaviour by updating gMe members from browser's devel-tool
+console.
 
-NOTE: It doesnt set any parameters other than temperature for now. However if someone wants they can update
-the js file as needed.
+NOTE: Given that the idea is for basic minimal testing, it doesnt bother with any model context length and
+culling of old messages from the chat by default. However by enabling the sliding window chat logic, a crude
+form of old messages culling can be achieved.
+
+NOTE: It doesnt set any parameters other than temperature and max_tokens for now. However if someone wants
+they can update the js file or equivalent member in gMe as needed.
 
 
 ## usage
@@ -43,11 +47,33 @@ next run this web front end in examples/server/public_simplechat
 ### using the front end
 
 Open this simple web front end from your local browser
+
 * http://127.0.0.1:PORT/index.html
 
 Once inside
+
 * Select between chat and completion mode. By default it is set to chat mode.
+
+* In completion mode
+  * logic by default doesnt insert any role specific "ROLE: " prefix wrt each role's message.
+    If the model requires any prefix wrt user role messages, then the end user has to
+    explicitly add the needed prefix, when they enter their chat message.
+    Similarly if the model requires any prefix to trigger assistant/ai-model response,
+    then the end user needs to enter the same.
+    This keeps the logic simple, while still giving flexibility to the end user to
+    manage any templating/tagging requirement wrt their messages to the model.
+  * the logic doesnt insert newline at the begining and end wrt the prompt message generated.
+    However if the chat being sent to /completions end point has more than one role's message,
+    then insert newline when moving from one role's message to the next role's message, so
+    that it can be clearly identified/distinguished.
+  * given that /completions endpoint normally doesnt add additional chat-templating of its
+    own, the above ensures that end user can create a custom single/multi message combo with
+    any tags/special-tokens related chat templating to test out model handshake. Or enduser
+    can use it just for normal completion related/based query.
+
 * If you want to provide a system prompt, then ideally enter it first, before entering any user query.
+  Normally Completion mode doesnt need system prompt, while Chat mode can generate better/interesting
+  responses with a suitable system prompt.
   * if chat.add_system_begin is used
     * you cant change the system prompt, after it is has been submitted once along with user query.
     * you cant set a system prompt, after you have submitted any user query
@@ -55,27 +81,121 @@ Once inside
     * one can change the system prompt any time during chat, by changing the contents of system prompt.
     * inturn the updated/changed system prompt will be inserted into the chat session.
     * this allows for the subsequent user chatting to be driven by the new system prompt set above.
+
 * Enter your query and either press enter or click on the submit button.
   If you want to insert enter (\n) as part of your chat/query to ai model, use shift+enter.
+
 * Wait for the logic to communicate with the server and get the response.
   * the user is not allowed to enter any fresh query during this time.
   * the user input box will be disabled and a working message will be shown in it.
+
 * just refresh the page, to reset wrt the chat history and or system prompt and start afresh.
+
 * Using NewChat one can start independent chat sessions.
   * two independent chat sessions are setup by default.
 
 
 ## Devel note
 
+### Reason behind this
+
+The idea is to be easy enough to use for basic purposes, while also being simple and easily discernable
+by developers who may not be from web frontend background (so inturn may not be familiar with template /
+end-use-specific-language-extensions driven flows) so that they can use it to explore/experiment things.
+
+And given that the idea is also to help explore/experiment for developers, some flexibility is provided
+to change behaviour easily using the devel-tools/console, for now. And skeletal logic has been implemented
+to explore some of the end points and ideas/implications around them.
+
+
+### General
+
+Me/gMe consolidates the settings which control the behaviour into one object.
+One can see the current settings, as well as change/update them using browsers devel-tool/console.
+
+  bCompletionFreshChatAlways - whether Completion mode collates complete/sliding-window history when
+  communicating with the server or only sends the latest user query/message.
+
+  bCompletionInsertStandardRolePrefix - whether Completion mode inserts role related prefix wrt the
+  messages that get inserted into prompt field wrt /Completion endpoint.
+
+  chatRequestOptions - maintains the list of options/fields to send along with chat request,
+  irrespective of whether /chat/completions or /completions endpoint.
+
+    If you want to add additional options/fields to send to the server/ai-model, and or
+    modify the existing options value or remove them, for now you can update this global var
+    using browser's development-tools/console.
+
+  iRecentUserMsgCnt - a simple minded SlidingWindow to limit context window load at Ai Model end.
+  This is disabled by default. However if enabled, then in addition to latest system message, only
+  the last/latest iRecentUserMsgCnt user messages after the latest system prompt and its responses
+  from the ai model will be sent to the ai-model, when querying for a new response. IE if enabled,
+  only user messages after the latest system message/prompt will be considered.
+
+    This specified sliding window user message count also includes the latest user query.
+    <0 : Send entire chat history to server
+     0 : Send only the system message if any to the server
+    >0 : Send the latest chat history from the latest system prompt, limited to specified cnt.
+
+
+By using gMe's iRecentUserMsgCnt and chatRequestOptions.max_tokens one can try to control the
+implications of loading of the ai-model's context window by chat history, wrt chat response to
+some extent in a simple crude way.
+
+
 Sometimes the browser may be stuborn with caching of the file, so your updates to html/css/js
 may not be visible. Also remember that just refreshing/reloading page in browser or for that
 matter clearing site data, dont directly override site caching in all cases. Worst case you may
 have to change port. Or in dev tools of browser, you may be able to disable caching fully.
 
+
 Concept of multiple chat sessions with different servers, as well as saving and restoring of
 those across browser usage sessions, can be woven around the SimpleChat/MultiChatUI class and
 its instances relatively easily, however given the current goal of keeping this simple, it has
 not been added, for now.
 
+
 By switching between chat.add_system_begin/anytime, one can control whether one can change
 the system prompt, anytime during the conversation or only at the beginning.
+
+
+read_json_early, is to experiment with reading json response data early on, if available,
+so that user can be shown generated data, as and when it is being generated, rather than
+at the end when full data is available.
+
+  the server flow doesnt seem to be sending back data early, atleast for request (inc options)
+  that is currently sent.
+
+  if able to read json data early on in future, as and when ai model is generating data, then
+  this helper needs to indirectly update the chat div with the recieved data, without waiting
+  for the overall data to be available.
+
+
+### Default setup
+
+By default things are setup to try and make the user experience a bit better, if possible.
+However a developer when testing the server of ai-model may want to change these value.
+
+Using iRecentUserMsgCnt reduce chat history context sent to the server/ai-model to be
+just the system-prompt, prev-user-request-and-ai-response and cur-user-request, instead of
+full chat history. This way if there is any response with garbage/repeatation, it doesnt
+mess with things beyond the next question/request/query, in some ways.
+
+Set max_tokens to 1024, so that a relatively large previous reponse doesnt eat up the space
+available wrt next query-response. However dont forget that the server when started should
+also be started with a model context size of 1k or more, to be on safe side.
+
+  The /completions endpoint of examples/server doesnt take max_tokens, instead it takes the
+  internal n_predict, for now add the same here on the client side, maybe later add max_tokens
+  to /completions endpoint handling code on server side.
+
+Frequency and presence penalty fields are set to 1.2 in the set of fields sent to server
+along with the user query. So that the model is partly set to try avoid repeating text in
+its response.
+
+A end-user can change these behaviour by editing gMe from browser's devel-tool/console.
+
+
+## At the end
+
+Also a thank you to all open source and open model developers, who strive for the common good.

examples/server/public_simplechat/simplechat.css

@@ -48,6 +48,13 @@ button {
     flex-direction: column;
 }
 
+.ul1 {
+    padding-inline-start: 2vw;
+}
+.ul2 {
+    padding-inline-start: 2vw;
+}
+
 * {
     margin: 0.6vmin;
 }

examples/server/public_simplechat/simplechat.js

@@ -14,23 +14,86 @@ class ApiEP {
 }
 
 let gUsageMsg = `
-    <p> Enter the system prompt above, before entering/submitting any user query.</p>
-    <p> Enter your text to the ai assistant below.</p>
-    <p> Use shift+enter for inserting enter.</p>
-    <p> Refresh the page to start over fresh.</p>
+    <p class="role-system">Usage</p>
+    <ul class="ul1">
+    <li> Set system prompt above, to try control ai response charactersitic, if model supports same.</li>
+        <ul class="ul2">
+        <li> Completion mode normally wont have a system prompt.</li>
+        </ul>
+    <li> Enter your query to ai assistant below.</li>
+        <ul class="ul2">
+        <li> Completion mode doesnt insert user/role: prefix implicitly.</li>
+        <li> Use shift+enter for inserting enter/newline.</li>
+        </ul>
+    <li> Default ContextWindow = [System, Last Query+Resp, Cur Query].</li>
+        <ul class="ul2">
+        <li> experiment iRecentUserMsgCnt, max_tokens, model ctxt window to expand</li>
+        </ul>
+    </ul>
 `;
 
+/** @typedef {{role: string, content: string}[]} ChatMessages */
+
 class SimpleChat {
 
     constructor() {
         /**
          * Maintain in a form suitable for common LLM web service chat/completions' messages entry
-         * @type {{role: string, content: string}[]}
+         * @type {ChatMessages}
          */
         this.xchat = [];
         this.iLastSys = -1;
     }
 
+    clear() {
+        this.xchat = [];
+        this.iLastSys = -1;
+    }
+
+    /**
+     * Recent chat messages.
+     * If iRecentUserMsgCnt < 0
+     *   Then return the full chat history
+     * Else
+     *   Return chat messages from latest going back till the last/latest system prompt.
+     *   While keeping track that the number of user queries/messages doesnt exceed iRecentUserMsgCnt.
+     * @param {number} iRecentUserMsgCnt
+     */
+    recent_chat(iRecentUserMsgCnt) {
+        if (iRecentUserMsgCnt < 0) {
+            return this.xchat;
+        }
+        if (iRecentUserMsgCnt == 0) {
+            console.warn("WARN:SimpleChat:SC:RecentChat:iRecentUsermsgCnt of 0 means no user message/query sent");
+        }
+        /** @type{ChatMessages} */
+        let rchat = [];
+        let sysMsg = this.get_system_latest();
+        if (sysMsg.length != 0) {
+            rchat.push({role: Roles.System, content: sysMsg});
+        }
+        let iUserCnt = 0;
+        let iStart = this.xchat.length;
+        for(let i=this.xchat.length-1; i > this.iLastSys; i--) {
+            if (iUserCnt >= iRecentUserMsgCnt) {
+                break;
+            }
+            let msg = this.xchat[i];
+            if (msg.role == Roles.User) {
+                iStart = i;
+                iUserCnt += 1;
+            }
+        }
+        for(let i = iStart; i < this.xchat.length; i++) {
+            let msg = this.xchat[i];
+            if (msg.role == Roles.System) {
+                continue;
+            }
+            rchat.push({role: msg.role, content: msg.content});
+        }
+        return rchat;
+    }
+
     /**
      * Add an entry into xchat
      * @param {string} role
@@ -57,7 +120,7 @@ class SimpleChat {
             div.replaceChildren();
         }
         let last = undefined;
-        for(const x of this.xchat) {
+        for(const x of this.recent_chat(gMe.iRecentUserMsgCnt)) {
             let entry = document.createElement("p");
             entry.className = `role-${x.role}`;
             entry.innerText = `${x.role}: ${x.content}`;
@@ -69,17 +132,21 @@ class SimpleChat {
         } else {
             if (bClear) {
                 div.innerHTML = gUsageMsg;
+                gMe.show_info(div);
             }
         }
     }
 
     /**
-     * Add needed fields wrt json object to be sent wrt LLM web services completions endpoint
+     * Add needed fields wrt json object to be sent wrt LLM web services completions endpoint.
+     * The needed fields/options are picked from a global object.
      * Convert the json into string.
      * @param {Object} obj
      */
     request_jsonstr(obj) {
-        obj["temperature"] = 0.7;
+        for(let k in gMe.chatRequestOptions) {
+            obj[k] = gMe.chatRequestOptions[k];
+        }
         return JSON.stringify(obj);
     }
 
@@ -88,18 +155,27 @@ class SimpleChat {
      */
     request_messages_jsonstr() {
         let req = {
-            messages: this.xchat,
+            messages: this.recent_chat(gMe.iRecentUserMsgCnt),
         }
         return this.request_jsonstr(req);
     }
 
     /**
      * Return a string form of json object suitable for /completions
+     * @param {boolean} bInsertStandardRolePrefix Insert "<THE_ROLE>: " as prefix wrt each role's message
      */
-    request_prompt_jsonstr() {
+    request_prompt_jsonstr(bInsertStandardRolePrefix) {
         let prompt = "";
-        for(const chat of this.xchat) {
-            prompt += `${chat.role}: ${chat.content}\n`;
+        let iCnt = 0;
+        for(const chat of this.recent_chat(gMe.iRecentUserMsgCnt)) {
+            iCnt += 1;
+            if (iCnt > 1) {
+                prompt += "\n";
+            }
+            if (bInsertStandardRolePrefix) {
+                prompt += `${chat.role}: `;
+            }
+            prompt += `${chat.content}`;
         }
         let req = {
             prompt: prompt,
@@ -171,7 +247,6 @@ let gChatURL = {
     'chat': `${gBaseURL}/chat/completions`,
     'completion': `${gBaseURL}/completions`,
 }
-const gbCompletionFreshChatAlways = true;
 
 
 /**
@@ -291,6 +366,8 @@ class MultiChatUI {
             // allow user to insert enter into their message using shift+enter.
             // while just pressing enter key will lead to submitting.
             if ((ev.key === "Enter") && (!ev.shiftKey)) {
+                let value = this.elInUser.value;
+                this.elInUser.value = value.substring(0,value.length-1);
                 this.elBtnUser.click();
                 ev.preventDefault();
             }
@@ -321,6 +398,29 @@ class MultiChatUI {
         }
     }
 
+    /**
+     * Try read json response early, if available.
+     * @param {Response} resp
+     */
+    async read_json_early(resp) {
+        if (!resp.body) {
+            throw Error("ERRR:SimpleChat:MCUI:ReadJsonEarly:No body...");
+        }
+        let tdUtf8 = new TextDecoder("utf-8");
+        let rr = resp.body.getReader();
+        let gotBody = "";
+        while(true) {
+            let { value: cur,  done: done} = await rr.read();
+            let curBody = tdUtf8.decode(cur);
+            console.debug("DBUG:SC:PART:", curBody);
+            gotBody += curBody;
+            if (done) {
+                break;
+            }
+        }
+        return JSON.parse(gotBody);
+    }
+
     /**
      * Handle user query submit request, wrt specified chat session.
      * @param {string} chatId
@@ -330,6 +430,14 @@ class MultiChatUI {
 
         let chat = this.simpleChats[chatId];
 
+        // In completion mode, if configured, clear any previous chat history.
+        // So if user wants to simulate a multi-chat based completion query,
+        // they will have to enter the full thing, as a suitable multiline
+        // user input/query.
+        if ((apiEP == ApiEP.Completion) && (gMe.bCompletionFreshChatAlways)) {
+            chat.clear();
+        }
+
         chat.add_system_anytime(this.elInSystem.value, chatId);
 
         let content = this.elInUser.value;
@@ -344,7 +452,7 @@ class MultiChatUI {
         if (apiEP == ApiEP.Chat) {
             theBody = chat.request_messages_jsonstr();
         } else {
-            theBody = chat.request_prompt_jsonstr();
+            theBody = chat.request_prompt_jsonstr(gMe.bCompletionInsertStandardRolePrefix);
         }
 
         this.elInUser.value = "working...";
@@ -359,6 +467,7 @@ class MultiChatUI {
         });
 
         let respBody = await resp.json();
+        //let respBody = await this.read_json_early(resp);
         console.debug(`DBUG:SimpleChat:MCUI:${chatId}:HandleUserSubmit:RespBody:${JSON.stringify(respBody)}`);
         let assistantMsg;
         if (apiEP == ApiEP.Chat) {
@@ -376,13 +485,6 @@ class MultiChatUI {
         } else {
             console.debug(`DBUG:SimpleChat:MCUI:HandleUserSubmit:ChatId has changed:[${chatId}] [${this.curChatId}]`);
         }
-        // Purposefully clear at end rather than begin of this function
-        // so that one can switch from chat to completion mode and sequece
-        // in a completion mode with multiple user-assistant chat data
-        // from before to be sent/occur once.
-        if ((apiEP == ApiEP.Completion) && (gbCompletionFreshChatAlways)) {
-            chat.xchat.length = 0;
-        }
         this.ui_reset_userinput();
     }
 
@@ -462,17 +564,66 @@ class MultiChatUI {
 }
 
 
-let gMuitChat;
-const gChatIds = [ "Default", "Other" ];
+class Me {
+
+    constructor() {
+        this.defaultChatIds = [ "Default", "Other" ];
+        this.multiChat = new MultiChatUI();
+        this.bCompletionFreshChatAlways = true;
+        this.bCompletionInsertStandardRolePrefix = false;
+        this.iRecentUserMsgCnt = 2;
+        // Add needed fields wrt json object to be sent wrt LLM web services completions endpoint.
+        this.chatRequestOptions = {
+            "temperature": 0.7,
+            "max_tokens": 1024,
+            "frequency_penalty": 1.2,
+            "presence_penalty": 1.2,
+            "n_predict": 1024
+        };
+    }
+
+    /**
+     * @param {HTMLDivElement} elDiv
+     */
+    show_info(elDiv) {
+
+        var p = document.createElement("p");
+        p.innerText = "Settings (devel-tools-console gMe)";
+        p.className = "role-system";
+        elDiv.appendChild(p);
+
+        var p = document.createElement("p");
+        p.innerText = `bCompletionFreshChatAlways:${this.bCompletionFreshChatAlways}`;
+        elDiv.appendChild(p);
+
+        p = document.createElement("p");
+        p.innerText = `bCompletionInsertStandardRolePrefix:${this.bCompletionInsertStandardRolePrefix}`;
+        elDiv.appendChild(p);
+
+        p = document.createElement("p");
+        p.innerText = `iRecentUserMsgCnt:${this.iRecentUserMsgCnt}`;
+        elDiv.appendChild(p);
+
+        p = document.createElement("p");
+        p.innerText = `chatRequestOptions:${JSON.stringify(this.chatRequestOptions)}`;
+        elDiv.appendChild(p);
+
+    }
+
+}
+
+
+/** @type {Me} */
+let gMe;
 
 function startme() {
     console.log("INFO:SimpleChat:StartMe:Starting...");
-    gMuitChat = new MultiChatUI();
-    for (let cid of gChatIds) {
-        gMuitChat.new_chat_session(cid);
+    gMe = new Me();
+    for (let cid of gMe.defaultChatIds) {
+        gMe.multiChat.new_chat_session(cid);
     }
-    gMuitChat.setup_ui(gChatIds[0]);
-    gMuitChat.show_sessions();
+    gMe.multiChat.setup_ui(gMe.defaultChatIds[0], true);
+    gMe.multiChat.show_sessions();
 }
 
 document.addEventListener("DOMContentLoaded", startme);

examples/sycl/win-build-sycl.bat

@@ -13,10 +13,10 @@ if %errorlevel% neq 0 goto ERROR
 
 ::  for FP16
 ::  faster for long-prompt inference
-::  cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
+::  cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
 
 ::  for FP32
-cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
+cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release
 if %errorlevel% neq 0 goto ERROR
 ::  build example/main only
 ::  make main

examples/tokenize/tokenize.cpp

@@ -3,40 +3,390 @@
 
 #include <cmath>
 #include <cstdio>
+#include <fstream>
 #include <string>
 #include <vector>
 
-int main(int argc, char ** argv) {
-    if (argc < 3 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH PROMPT [--ids]\n" , argv[0]);
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#include <shellapi.h>   // For CommandLineToArgvW
+#endif
+
+static void print_usage_information(const char * argv0, FILE * stream) {
+    fprintf(stream, "usage: %s [options]\n\n", argv0);
+    fprintf(stream, "The tokenize program tokenizes a prompt using a given model,\n");
+    fprintf(stream, "and prints the resulting tokens to standard output.\n\n");
+    fprintf(stream, "It needs a model file, a prompt, and optionally other flags\n");
+    fprintf(stream, "to control the behavior of the tokenizer.\n\n");
+    fprintf(stream, "    The possible options are:\n");
+    fprintf(stream, "\n");
+    fprintf(stream, "    -h, --help                           print this help and exit\n");
+    fprintf(stream, "    -m MODEL_PATH, --model MODEL_PATH    path to model.\n");
+    fprintf(stream, "    --ids                                if given, only print numerical token IDs, and not token strings.\n");
+    fprintf(stream, "                                         The output format looks like [1, 2, 3], i.e. parseable by Python.\n");
+    fprintf(stream, "    -f PROMPT_FNAME, --file PROMPT_FNAME read prompt from a file.\n");
+    fprintf(stream, "    -p PROMPT, --prompt PROMPT           read prompt from the argument.\n");
+    fprintf(stream, "    --stdin                              read prompt from standard input.\n");
+    fprintf(stream, "    --no-bos                             do not ever add a BOS token to the prompt, even if normally the model uses a BOS token.\n");
+    fprintf(stream, "    --log-disable                        disable logs. Makes stderr quiet when loading the model.\n");
+}
+
+static void llama_log_callback_null(ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) text;
+    (void) user_data;
+}
+
+static std::string read_prompt_from_file(const char * filepath, bool & success) {
+    success = false;
+
+    std::ifstream in(filepath, std::ios::binary);
+    if (!in) {
+        fprintf(stderr, "%s: could not open file '%s' for reading: %s\n", __func__, filepath, strerror(errno));
+        return std::string();
+    }
+    // do not assume the file is seekable (e.g. /dev/stdin)
+    std::stringstream buffer;
+    buffer << in.rdbuf();
+    if (in.fail()) {
+        fprintf(stderr, "%s: could not read the entire file '%s': %s\n", __func__, filepath, strerror(errno));
+        return std::string();
+    }
+
+    success = true;
+    return buffer.str();
+}
+
+//
+// Function: ingest_args(...) -> vector<string>
+//
+//  Takes argc and argv arguments, and converts them to a vector of UTF-8 encoded
+//  strings, as an STL vector<string>.
+//
+//  In particular, it handles character encoding shenanigans on Windows.
+//
+// Note: raw_argc and raw_argv are not actually read at all on Windows.
+//       On Windows we call GetCommandLineW to get the arguments in wchar_t
+//       format, ignoring the regular argc/argv arguments to main().
+//
+// TODO: potential opportunity to roll common stuff into common/console.cpp
+//       in relation to Windows wchar_t shenanigans.
+static std::vector<std::string> ingest_args(int raw_argc, char ** raw_argv) {
+    std::vector<std::string> argv;
+
+    // Handle Windows, if given non-ASCII arguments.
+    // We convert wchar_t arguments into UTF-8 char* on this platform.
+    // Lets you invoke 'tokenize' on Windows cmd.exe with non-ASCII characters
+    // without throwing tantrums.
+#if defined(_WIN32)
+    int argc;
+    const LPWSTR cmdline_wargv = GetCommandLineW();
+    LPWSTR * wargv = CommandLineToArgvW(cmdline_wargv, &argc);
+
+    // silence unused arg warnings
+    (void) raw_argc;
+    (void) raw_argv;
+
+    for (int i = 0; i < argc; ++i) {
+        int length_needed = WideCharToMultiByte(CP_UTF8, 0, wargv[i], wcslen(wargv[i]), 0, 0, NULL, NULL);
+        char * output_buf = (char *) calloc(length_needed+1, sizeof(char));
+        GGML_ASSERT(output_buf);
+
+        WideCharToMultiByte(CP_UTF8, 0, wargv[i], wcslen(wargv[i]), output_buf, length_needed, NULL, NULL);
+        output_buf[length_needed] = '\0';
+
+        argv.push_back(output_buf);
+        free(output_buf);
+    }
+
+    LocalFree((HLOCAL) wargv);
+#else
+    int argc = raw_argc;
+    for (int i = 0; i < argc; ++i) {
+        argv.push_back(raw_argv[i]);
+    }
+#endif
+
+    GGML_ASSERT((unsigned int) argc == argv.size());
+
+    return argv;
+}
+
+//
+// Function: write_utf8_cstr_to_stdout(const char *) -> <writes to stdout>
+//
+// writes a string to standard output; taking into account that on Windows
+// to display correctly you have to use special handling. Works even if the
+// user has not set a unicode code page on a Windows cmd.exe.
+//
+// In case of invalid UTF-8, invalid_utf8 is set to true on Windows, and something
+// a human-readable is written instead.
+//
+// On non-Windows systems, simply printfs() the string.
+static void write_utf8_cstr_to_stdout(const char * str, bool & invalid_utf8) {
+        invalid_utf8 = false;
+
+#if defined(_WIN32)
+        // Are we in a console?
+        HANDLE hConsole = GetStdHandle(STD_OUTPUT_HANDLE);
+        DWORD dwMode = 0;
+
+        // According to Microsoft docs:
+        // "WriteConsole fails if it is used with a standard handle that is redirected to a file."
+        // Also according to the docs, you can use GetConsoleMode to check for that.
+        if (hConsole == INVALID_HANDLE_VALUE || !GetConsoleMode(hConsole, &dwMode)) {
+            printf("%s", str);
+            return;
+        }
+
+        // MultiByteToWideChar reports an error if str is empty, don't report
+        // them as invalid_utf8.
+        if (*str == 0) {
+            return;
+        }
+        int length_needed = MultiByteToWideChar(CP_UTF8, MB_ERR_INVALID_CHARS, str, strlen(str), NULL, 0);
+        if (length_needed == 0) {
+            DWORD err = GetLastError();
+            if (err == ERROR_NO_UNICODE_TRANSLATION) {
+                invalid_utf8 = true;
+                int len = strlen(str);
+                printf("<");
+                for (int i = 0; i < len; ++i) {
+                    if (i > 0) {
+                        printf(" ");
+                    }
+                    printf("%02x", (uint8_t) str[i]);
+                }
+                printf(">");
+                return;
+            }
+            GGML_ASSERT(false && "MultiByteToWideChar() failed in an unexpected way.");
+        }
+
+        LPWSTR wstr = (LPWSTR) calloc(length_needed+1, sizeof(*wstr));
+        GGML_ASSERT(wstr);
+
+        MultiByteToWideChar(CP_UTF8, 0, str, strlen(str), wstr, length_needed);
+        WriteConsoleW(hConsole, wstr, length_needed, NULL, NULL);
+
+        free(wstr);
+#else
+        // TODO: reporting invalid_utf8 would be useful on non-Windows too.
+        // printf will silently just write bad unicode.
+        printf("%s", str);
+#endif
+}
+
+int main(int raw_argc, char ** raw_argv) {
+    const std::vector<std::string> argv = ingest_args(raw_argc, raw_argv);
+    const int argc = argv.size();
+
+    if (argc <= 1) {
+        print_usage_information(argv[0].c_str(), stderr);
         return 1;
     }
 
-    const char * model_path = argv[1];
-    const char * prompt     = argv[2];
+    //////
+    // Read out all the command line arguments.
+    //////
 
-    const bool printing_ids = argc > 3 && std::string(argv[3]) == "--ids";
+    // variables where to put any arguments we see.
+    bool printing_ids = false;
+    bool no_bos = false;
+    bool disable_logging = false;
+    const char * model_path = NULL;
+    const char * prompt_path = NULL;
+    const char * prompt_arg = NULL;
+
+    // track which arguments were explicitly given
+    // used for sanity checking down the line
+    bool model_path_set = false;
+    bool prompt_path_set = false;
+    bool prompt_set = false;
+    bool stdin_set = false;
+
+    int iarg = 1;
+    for (; iarg < argc; ++iarg) {
+        std::string arg{argv[iarg]};
+        if (arg == "-h" || arg == "--help") {
+            print_usage_information(argv[0].c_str(), stdout);
+            return 0;
+        }
+        else if (arg == "--ids") {
+            printing_ids = true;
+        }
+        else if (arg == "-m" || arg == "--model") {
+            if (model_path_set) {
+                fprintf(stderr, "Error: -m or --model specified multiple times.\n");
+                return 1;
+            }
+            model_path = argv[++iarg].c_str();
+            model_path_set = true;
+        }
+        else if (arg == "--no-bos") {
+            no_bos = true;
+        }
+        else if (arg == "-p" || arg == "--prompt") {
+            if (prompt_set) {
+                fprintf(stderr, "Error: -p or --prompt specified multiple times.\n");
+                return 1;
+            }
+            prompt_arg = argv[++iarg].c_str();
+            prompt_set = true;
+        }
+        else if (arg == "-f" || arg == "--file") {
+            if (prompt_path_set) {
+                fprintf(stderr, "Error: -f or --file specified multiple times.\n");
+                return 1;
+            }
+            prompt_path = argv[++iarg].c_str();
+            prompt_path_set = true;
+        }
+        else if (arg == "--stdin") {
+            stdin_set = true;
+        }
+        else if (arg == "--log-disable") {
+            disable_logging = true;
+        }
+        else {
+            fprintf(stderr, "Error: unknown option '%s'\n", argv[iarg].c_str());
+            return 1;
+        }
+    }
+
+    //////
+    // Sanity check the command line arguments.
+    //////
+
+    // Check that we have the required stuff set.
+    if (model_path_set && model_path == NULL) {
+        fprintf(stderr, "Error: --model requires an argument.\n");
+        return 1;
+    }
+    if (!model_path_set) {
+        fprintf(stderr, "Error: must specify --model.\n");
+        return 1;
+    }
+    if (prompt_path_set && prompt_path == NULL) {
+        fprintf(stderr, "Error: --file requires an argument.\n");
+        return 1;
+    }
+    if (prompt_set && prompt_arg == NULL) {
+        fprintf(stderr, "Error: --prompt requires an argument.\n");
+        return 1;
+    }
+    const int prompts_set = !!(prompt_path_set) + !!(prompt_set) + !!(stdin_set);
+    if (prompts_set > 1) {
+        fprintf(stderr, "Error: --stdin, --file and --prompt are mutually exclusive.\n");
+        return 1;
+    }
+    // Must have some prompt.
+    if (prompts_set == 0) {
+        fprintf(stderr, "Error: must specify one of: --stdin, --file or --prompt.\n");
+        return 1;
+    }
+
+    GGML_ASSERT(model_path);
+    GGML_ASSERT(prompt_path || prompt_arg || stdin_set);
+
+    //////
+    // Figure out where will the prompt come from.
+    //////
+
+    std::string prompt;
+    if (prompt_path_set) {
+        bool success = false;
+        prompt = read_prompt_from_file(prompt_path, success);
+        if (!success) {
+            return 1;
+        }
+    } else if (prompt_set) {
+        prompt = prompt_arg;
+    } else {
+        GGML_ASSERT(stdin_set);
+        // we read stdin *after* loading model (early exit if model cannot
+        // be loaded, which can be a nicer user experience)
+    }
+
+    //////
+    // Start actually doing the tokenizing stuff.
+    //////
+
+#ifdef LOG_DISABLE_LOGS
+    disable_logging = true;
+#endif
+
+    if (disable_logging) {
+        llama_log_set(llama_log_callback_null, NULL);
+    }
 
     llama_backend_init();
 
     llama_model_params model_params = llama_model_default_params();
     model_params.vocab_only = true;
     llama_model * model = llama_load_model_from_file(model_path, model_params);
+    if (!model) {
+        fprintf(stderr, "Error: could not load model from file '%s'.\n", model_path);
+        return 1;
+    }
 
     llama_context_params ctx_params = llama_context_default_params();
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
+    if (!ctx) {
+        fprintf(stderr, "Error: could not create context.\n");
+        return 1;
+    }
+
+    // read entire prompt from stdin?
+    if (stdin_set) {
+        GGML_ASSERT(!prompt_path_set && !prompt_set);
+
+        std::stringstream stdin_buffer;
+        stdin_buffer << std::cin.rdbuf();
+        if (std::cin.fail()) {
+            fprintf(stderr, "Error: could not read the entire standard input.\n");
+            return 1;
+        }
+
+        prompt = stdin_buffer.str();
+    }
+
+    const bool model_wants_add_bos = llama_should_add_bos_token(model);
+    const bool add_bos = model_wants_add_bos && !no_bos;
 
     std::vector<llama_token> tokens;
+    tokens = ::llama_tokenize(model, prompt, add_bos, true);
 
-    tokens = ::llama_tokenize(model, prompt, true, true);
+    if (printing_ids) {
+        printf("[");
+    }
 
     for (int i = 0; i < (int) tokens.size(); i++) {
         if (printing_ids) {
-            printf("%d\n", tokens[i]);
+            if (i > 0) {
+                printf(", ");
+            }
+            printf("%d", tokens[i]);
         } else {
-            printf("%6d -> '%s'\n", tokens[i], llama_token_to_piece(ctx, tokens[i]).c_str());
+            bool invalid_utf8 = false;
+            printf("%6d -> '", tokens[i]);
+            write_utf8_cstr_to_stdout(llama_token_to_piece(ctx, tokens[i]).c_str(), invalid_utf8);
+            if (invalid_utf8) {
+                printf("' (utf-8 decode failure)\n");
+            } else {
+                printf("'\n");
+            }
         }
     }
 
+    if (printing_ids) {
+        printf("]\n");
+    }
+
+    // silence valgrind
+    llama_free(ctx);
+    llama_free_model(model);
+
     return 0;
 }

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -341,7 +341,8 @@ static struct ggml_tensor * llama_build_train_graphs(
         struct ggml_tensor * t15 = ggml_permute      (ctx, t12, 0, 3, 1, 2);                        set_name(t15, "t15");     assert_shape_4d(t15, N, n_embd/n_head, n_head, n_batch);
         struct ggml_tensor * t16;
         if (enable_flash_attn) {
-            t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                        set_name(t16, "t16");     assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
+            GGML_ASSERT(false && "TODO: ggml_flash_attn_ext() not yet supported");
+            //t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                        set_name(t16, "t16");     assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
         } else {
             struct ggml_tensor * t16_0 = ggml_mul_mat              (ctx, t14, t13);                 set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
             struct ggml_tensor * t16_1 = ggml_scale_inplace        (ctx, t16_0, kv_scale);          set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);

flake.lock

@@ -5,11 +5,11 @@
         "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1714641030,
-        "narHash": "sha256-yzcRNDoyVP7+SCNX0wmuDju1NUCt8Dz9+lyUXEI0dbI=",
+        "lastModified": 1715865404,
+        "narHash": "sha256-/GJvTdTpuDjNn84j82cU6bXztE0MSkdnTWClUCRub78=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "e5d10a24b66c3ea8f150e47dfdb0416ab7c3390e",
+        "rev": "8dc45382d5206bd292f9c2768b8058a8fd8311d9",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1714635257,
-        "narHash": "sha256-4cPymbty65RvF1DWQfc+Bc8B233A1BWxJnNULJKQ1EY=",
+        "lastModified": 1716509168,
+        "narHash": "sha256-4zSIhSRRIoEBwjbPm3YiGtbd8HDWzFxJjw5DYSDy1n8=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "63c3a29ca82437c87573e4c6919b09a24ea61b0f",
+        "rev": "bfb7a882678e518398ce9a31a881538679f6f092",
         "type": "github"
       },
       "original": {

ggml-common.h

@@ -65,13 +65,8 @@ typedef sycl::half2 ggml_half2;
 // QK = number of values after dequantization
 // QK_K = super-block size
 
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
 #define QK_K 256
 #define K_SCALE_SIZE 12
-#endif // GGML_QKK_64
 
 #if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL)
 // QR = QK / number of values before dequantization
@@ -131,13 +126,8 @@ typedef sycl::half2 ggml_half2;
 #define QI4_NL (QK4_NL / (4*QR4_NL))
 #define QR4_NL 2
 
-#if QK_K == 64
-#define QI4_XS QI4_NL
-#define QR4_XS QR4_NL
-#else
 #define QI4_XS (QK_K / (4*QR4_XS))
 #define QR4_XS 8
-#endif
 
 #endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
 
@@ -228,15 +218,6 @@ static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wro
 // weight is represented as x = a * q
 // 16 blocks of 16 elements each
 // Effectively 3.4375 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    uint8_t hmask[QK_K/8]; // quants - high bit
-    uint8_t qs[QK_K/4];    // quants - low 2 bits
-    uint8_t scales[2];
-    ggml_half d;           // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
-#else
 typedef struct {
     uint8_t hmask[QK_K/8]; // quants - high bit
     uint8_t qs[QK_K/4];    // quants - low 2 bits
@@ -244,20 +225,11 @@ typedef struct {
     ggml_half d;           // super-block scale
 } block_q3_K;
 static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
-#endif
 
 // 4-bit quantization
 // 8 blocks of 32 elements each
 // weight is represented as x = a * q + b
 // Effectively 4.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_half d[2];     // super-block scales/mins
-    uint8_t scales[2];  // 4-bit block scales/mins
-    uint8_t qs[QK_K/2]; // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
 typedef struct {
     union {
         struct {
@@ -270,21 +242,11 @@ typedef struct {
     uint8_t qs[QK_K/2];           // 4--bit quants
 } block_q4_K;
 static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
-#endif
 
 // 5-bit quantization
 // 8 blocks of 32 elements each
 // weight is represented as x = a * q + b
 // Effectively 5.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_half d;             // super-block scale
-    int8_t  scales[QK_K/16]; // 8-bit block scales
-    uint8_t qh[QK_K/8];      // quants, high bit
-    uint8_t qs[QK_K/2];      // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_half) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
 typedef struct {
     union {
         struct {
@@ -298,7 +260,6 @@ typedef struct {
     uint8_t qs[QK_K/2];           // quants, low 4 bits
 } block_q5_K;
 static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif
 
 // 6-bit quantization
 // weight is represented as x = a * q
@@ -356,11 +317,7 @@ typedef struct {
 static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
 
 // 3.4375 bpw
-#if QK_K == 64
-#define IQ3S_N_SCALE 2
-#else
 #define IQ3S_N_SCALE QK_K/64
-#endif
 typedef struct {
     ggml_half d;
     uint8_t qs[QK_K/4];
@@ -381,16 +338,9 @@ static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wron
 typedef struct {
     uint8_t  qs[QK_K/8];      // grid index, low 8 bits
     uint8_t  qh[QK_K/16];     // grid index, high 3 bits + grid shift bit (for two groups of 8)
-#if QK_K == 64
-    ggml_half d;
-#endif
     uint8_t  scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64)
 } block_iq1_m;
-#if QK_K == 64
-static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32 + sizeof(ggml_half), "wrong iq1_m block size/padding");
-#else
 static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding");
-#endif
 
 // Used by IQ1_M quants
 typedef union {
@@ -406,9 +356,6 @@ typedef struct {
 } block_iq4_nl;
 static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding");
 
-#if QK_K == 64
-#define block_iq4_xs block_iq4_nl
-#else
 typedef struct {
     ggml_half d;
     uint16_t scales_h;
@@ -416,7 +363,6 @@ typedef struct {
     uint8_t  qs[QK_K/2];
 } block_iq4_xs;
 static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
-#endif
 
 #endif // GGML_COMMON_DECL
 #endif // GGML_COMMON_DECL

ggml-cuda.cu

@@ -119,6 +119,20 @@ int ggml_cuda_get_device() {
     return id;
 }
 
+static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
+    ggml_cuda_set_device(device);
+#if defined(GGML_USE_HIPBLAS) && defined(GGML_HIP_UMA)
+    auto res = hipMallocManaged(ptr, size);
+    if (res == hipSuccess) {
+        // if error we "need" to know why...
+        CUDA_CHECK(hipMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
+    }
+    return res;
+#else
+    return cudaMalloc(ptr, size);
+#endif
+}
+
 static ggml_cuda_device_info ggml_cuda_init() {
 #ifdef __HIP_PLATFORM_AMD__
     // Workaround for a rocBLAS bug when using multiple graphics cards:
@@ -271,7 +285,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
         size_t look_ahead_size = (size_t) (1.05 * size);
         look_ahead_size = 256 * ((look_ahead_size + 255)/256);
         ggml_cuda_set_device(device);
-        CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size));
+        CUDA_CHECK(ggml_cuda_device_malloc(&ptr, look_ahead_size, device));
         *actual_size = look_ahead_size;
         pool_size += look_ahead_size;
 #ifdef DEBUG_CUDA_MALLOC
@@ -537,7 +551,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffe
     size = std::max(size, (size_t)1); // cudaMalloc returns null for size 0
 
     void * dev_ptr;
-    cudaError_t err = cudaMalloc(&dev_ptr, size);
+    cudaError_t err = ggml_cuda_device_malloc(&dev_ptr, size, buft_ctx->device);
     if (err != cudaSuccess) {
         // clear the error
         cudaGetLastError();
@@ -798,7 +812,7 @@ GGML_CALL static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_bu
         // currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first
         ggml_cuda_set_device(id);
         char * buf;
-        CUDA_CHECK(cudaMalloc(&buf, size));
+        CUDA_CHECK(ggml_cuda_device_malloc((void**)&buf, size, id));
 
         // set padding to 0 to avoid possible NaN values
         if (size > original_size) {
@@ -2510,9 +2524,9 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
 
     bool use_cuda_graph = true;
     bool cuda_graph_update_required = false;
-    // pointer to CUDA cpy kernel, which is required to identify
+    // vector of pointers to CUDA cpy kernels, which are required to identify
     // kernel parameters which need updated in the graph for each token
-    void * ggml_cuda_cpy_fn_ptr = nullptr;
+    std::vector<void *> ggml_cuda_cpy_fn_ptrs;
 
     if (cuda_ctx->cuda_graph->graph == nullptr) {
         if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) {
@@ -2588,9 +2602,10 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
             if (node->op == GGML_OP_CPY) {
                 // store the copy op parameter which changes with each token.
                 cuda_ctx->cuda_graph->updated_kernel_arg.push_back((char **) &(node->src[1]->data));
-                if (ggml_cuda_cpy_fn_ptr == nullptr) {
-                    // store a pointer to the copy op CUDA kernel to identify it later
-                    ggml_cuda_cpy_fn_ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
+                // store a pointer to each copy op CUDA kernel to identify it later
+                void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
+                if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) {
+                    ggml_cuda_cpy_fn_ptrs.push_back(ptr);
                 }
             }
 
@@ -2720,7 +2735,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
         if (!cuda_graph_update_required) { // on update steps, the live parameters will already be captured
             int k = 0;
             for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
-                if (cuda_ctx->cuda_graph->params[i].func == ggml_cuda_cpy_fn_ptr) {
+                if(count(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), cuda_ctx->cuda_graph->params[i].func) > 0) {
                     char ** updated_kernel_arg_ptr = cuda_ctx->cuda_graph->updated_kernel_arg.at(k++);
                     cuda_ctx->cuda_graph->params[i].kernelParams[1] = updated_kernel_arg_ptr;
                     CUDA_CHECK(cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]));

ggml-cuda/common.cuh

@@ -79,13 +79,8 @@
 #define cudaHostRegisterReadOnly hipHostRegisterReadOnly
 #define cudaHostUnregister hipHostUnregister
 #define cudaLaunchHostFunc hipLaunchHostFunc
-#ifdef GGML_HIP_UMA
-#define cudaMalloc hipMallocManaged
-#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size)
-#else
 #define cudaMalloc hipMalloc
 #define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
-#endif
 #define cudaMemcpy hipMemcpy
 #define cudaMemcpyAsync hipMemcpyAsync
 #define cudaMemcpyPeerAsync hipMemcpyPeerAsync

ggml-cuda/convert.cu

@@ -131,7 +131,6 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
     const block_q2_K * x = (const block_q2_K *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t n   = tid/32;
     const int64_t l   = tid - 32*n;
     const int64_t is  = 8*n + l/16;
@@ -145,17 +144,6 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
     y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
     y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
     y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
-#else
-    const int64_t is = tid/16;  // 0 or 1
-    const int64_t il = tid%16;  // 0...15
-    const uint8_t q = x[i].qs[il] >> (2*is);
-    dst_t * y = yy + i*QK_K + 16*is + il;
-    float dall = __low2half(x[i].dm);
-    float dmin = __high2half(x[i].dm);
-    y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
-    y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
-#endif
-
 }
 
 template<typename dst_t>
@@ -164,7 +152,6 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t
     const int64_t i = blockIdx.x;
     const block_q3_K * x = (const block_q3_K *) vx;
 
-#if QK_K == 256
     const int64_t r = threadIdx.x/4;
     const int64_t tid = r/2;
     const int64_t is0 = r%2;
@@ -188,31 +175,8 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t
     const uint8_t * hm = x[i].hmask;
 
     for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
-#else
-    const int64_t tid = threadIdx.x;
-    const int64_t is  = tid/16;  // 0 or 1
-    const int64_t il  = tid%16;  // 0...15
-    const int64_t im  = il/8;    // 0...1
-    const int64_t in  = il%8;    // 0...7
-
-    dst_t * y = yy + i*QK_K + 16*is + il;
-
-    const uint8_t q = x[i].qs[il] >> (2*is);
-    const uint8_t h = x[i].hmask[in] >> (2*is + im);
-    const float   d = (float)x[i].d;
-
-    if (is == 0) {
-        y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
-        y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
-    } else {
-        y[ 0] = d * ((x[i].scales[0] >>  4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
-        y[32] = d * ((x[i].scales[1] >>  4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
-    }
-#endif
-
 }
 
-#if QK_K == 256
 static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
     if (j < 4) {
         d = q[j] & 63; m = q[j + 4] & 63;
@@ -221,7 +185,6 @@ static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t
         m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
     }
 }
-#endif
 
 template<typename dst_t>
 static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
@@ -229,7 +192,6 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t
 
     const int64_t i = blockIdx.x;
 
-#if QK_K == 256
     // assume 32 threads
     const int64_t tid = threadIdx.x;
     const int64_t il  = tid/8;
@@ -253,15 +215,6 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t
         y[l + 0] = d1 * (q[l] & 0xF) - m1;
         y[l +32] = d2 * (q[l] >>  4) - m2;
     }
-#else
-    const int64_t tid = threadIdx.x;
-    const uint8_t * q = x[i].qs;
-    dst_t * y = yy + i*QK_K;
-    const float d = (float)x[i].dm[0];
-    const float m = (float)x[i].dm[1];
-    y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
-    y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >>  4) - m * (x[i].scales[1] >> 4);
-#endif
 }
 
 template<typename dst_t>
@@ -270,7 +223,6 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t
 
     const int64_t i = blockIdx.x;
 
-#if QK_K == 256
     // assume 64 threads - this is very slightly better than the one below
     const int64_t tid = threadIdx.x;
     const int64_t il  = tid/16;   // il is in 0...3
@@ -297,18 +249,6 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t
     hm <<= 1;
     y[32] = d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2;
     y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
-#else
-    const int64_t tid = threadIdx.x;
-    const uint8_t q = x[i].qs[tid];
-    const int64_t im = tid/8;  // 0...3
-    const int64_t in = tid%8;  // 0...7
-    const int64_t is = tid/16; // 0 or 1
-    const uint8_t h = x[i].qh[in] >> im;
-    const float d = x[i].d;
-    dst_t * y = yy + i*QK_K + tid;
-    y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16));
-    y[32] = d * x[i].scales[is+2] * ((q >>  4) - ((h >> 4) & 1 ? 0 : 16));
-#endif
 }
 
 template<typename dst_t>
@@ -316,7 +256,6 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
     const block_q6_K * x = (const block_q6_K *) vx;
 
     const int64_t i = blockIdx.x;
-#if QK_K == 256
 
     // assume 64 threads - this is very slightly better than the one below
     const int64_t tid = threadIdx.x;
@@ -336,24 +275,6 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
     y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
     y[64] = d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
     y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
-#else
-
-    // assume 32 threads
-    const int64_t tid = threadIdx.x;
-    const int64_t ip  = tid/16;         // 0 or 1
-    const int64_t il  = tid - 16*ip;    // 0...15
-
-    dst_t * y = yy + i*QK_K + 16*ip + il;
-
-    const float d = x[i].d;
-
-    const uint8_t   ql = x[i].ql[16*ip + il];
-    const uint8_t   qh = x[i].qh[il] >> (2*ip);
-    const int8_t  * sc = x[i].scales;
-
-    y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
-    y[32] = d * sc[ip+2] * ((int8_t)((ql  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
-#endif
 }
 
 template<typename dst_t>
@@ -363,7 +284,6 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
     const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -374,10 +294,6 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
     const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
     const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
     for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -387,7 +303,6 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
     const block_iq2_xs * x = (const block_iq2_xs *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -396,10 +311,6 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
     const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
     const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
     for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -409,7 +320,6 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
     const block_iq2_s * x = (const block_iq2_s *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -417,10 +327,6 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
     const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
     const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
     for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -430,7 +336,6 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
     const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -445,10 +350,6 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
         y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
         y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
     }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -458,7 +359,6 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
     const block_iq3_s * x = (const block_iq3_s *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -471,10 +371,6 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
         y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
         y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
     }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -484,7 +380,6 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
     const block_iq1_s * x = (const block_iq1_s  *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -497,10 +392,6 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
     for (int j = 0; j < 8; ++j) {
         y[j] = d * (q[j] + delta);
     }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -510,7 +401,6 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
     const block_iq1_m * x = (const block_iq1_m  *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -527,13 +417,8 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
     for (int j = 0; j < 8; ++j) {
         y[j] = d * (q[j] + delta);
     }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
-
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
@@ -550,10 +435,8 @@ static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst
         y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
         y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
     }
-
 }
 
-#if QK_K != 64
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
     const int64_t i   = blockIdx.x;
@@ -570,7 +453,6 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst
         y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
     }
 }
-#endif
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
@@ -592,21 +474,13 @@ static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half *
 template<typename dst_t>
 static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q2_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template<typename dst_t>
 static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q3_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template<typename dst_t>
@@ -632,21 +506,13 @@ static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int64_t k
 template<typename dst_t>
 static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q5_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template<typename dst_t>
 static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q6_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template<typename dst_t>
@@ -700,11 +566,7 @@ static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int64_t
 template<typename dst_t>
 static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = (k + QK_K - 1) / QK_K;
-#if QK_K == 64
-    dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
-#else
     dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template <typename src_t, typename dst_t>

ggml-cuda/dmmv.cu

@@ -22,7 +22,6 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
 
     float tmp = 0; // partial sum for thread in warp
 
-#if QK_K == 256
     const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...15
     const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
 
@@ -71,37 +70,6 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
         tmp += dall * sum1 - dmin * sum2;
 
     }
-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
-    const int offset = tid * K_QUANTS_PER_ITERATION;
-
-    uint32_t uaux[2];
-    const uint8_t * d = (const uint8_t *)uaux;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + offset;
-        const uint8_t * q = x[i].qs + offset;
-        const uint32_t * s = (const uint32_t *)x[i].scales;
-
-        uaux[0] = s[0] & 0x0f0f0f0f;
-        uaux[1] = (s[0] >> 4) & 0x0f0f0f0f;
-
-        const float2 dall = __half22float2(x[i].dm);
-
-        float sum1 = 0, sum2 = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            const uint8_t ql = q[l];
-            sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3)
-                  + y[l+16] * d[1] * ((ql >> 2) & 3)
-                  + y[l+32] * d[2] * ((ql >> 4) & 3)
-                  + y[l+48] * d[3] * ((ql >> 6) & 3);
-            sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7];
-        }
-        tmp += dall.x * sum1 - dall.y * sum2;
-    }
-#endif
 
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
@@ -123,8 +91,6 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx,
 
     float tmp = 0; // partial sum for thread in warp
 
-#if QK_K == 256
-
     const uint16_t kmask1 = 0x0303;
     const uint16_t kmask2 = 0x0f0f;
 
@@ -175,34 +141,6 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx,
         tmp += d * sum;
 
     }
-#else
-
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
-    const int offset = tid * K_QUANTS_PER_ITERATION;         // 0...15 or 0...14
-    const int in = offset/8;                                 // 0 or 1
-    const int im = offset%8;                                 // 0...7
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + offset;
-        const uint8_t * q = x[i].qs + offset;
-        const uint8_t * s = x[i].scales;
-
-        const float dall = (float)x[i].d;
-
-        float sum = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            const uint8_t hl = x[i].hmask[im+l] >> in;
-            const uint8_t ql = q[l];
-            sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4))
-                 + y[l+16] * dall * ((s[0] >>  4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4))
-                 + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4))
-                 + y[l+48] * dall * ((s[1] >>  4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4));
-        }
-        tmp += sum;
-    }
-#endif
 
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
@@ -221,7 +159,6 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
 
     const block_q4_K * x = (const block_q4_K *)vx + ib0;
 
-#if QK_K == 256
     const uint16_t kmask1 = 0x3f3f;
     const uint16_t kmask2 = 0x0f0f;
     const uint16_t kmask3 = 0xc0c0;
@@ -306,36 +243,6 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
 #endif
 
     }
-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
-
-    const int step = tid * K_QUANTS_PER_ITERATION;
-
-    uint16_t aux16[2];
-    const uint8_t * s = (const uint8_t *)aux16;
-
-    float tmp = 0;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-        const uint8_t * q = x[i].qs + step;
-        const float   * y = yy + i*QK_K + step;
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux16[0] = a[0] & 0x0f0f;
-        aux16[1] = (a[0] >> 4) & 0x0f0f;
-        const float d = (float)x[i].dm[0];
-        const float m = (float)x[i].dm[1];
-        float sum = 0.f;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
-                 + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2])
-                 + y[j+32] * (d * s[1] * (q[j+ 0] >>  4) - m * s[3])
-                 + y[j+48] * (d * s[1] * (q[j+16] >>  4) - m * s[3]);
-        }
-        tmp += sum;
-    }
-
-#endif
 
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
@@ -355,7 +262,6 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
 
     float tmp = 0; // partial sum for thread in warp
 
-#if QK_K == 256
     const uint16_t kmask1 = 0x3f3f;
     const uint16_t kmask2 = 0x0f0f;
     const uint16_t kmask3 = 0xc0c0;
@@ -426,30 +332,6 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
         tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
     }
 
-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
-    const int step = tid * K_QUANTS_PER_ITERATION;
-    const int im = step/8;
-    const int in = step%8;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-        const uint8_t * q = x[i].qs + step;
-        const int8_t  * s = x[i].scales;
-        const float   * y = yy + i*QK_K + step;
-        const float     d = x[i].d;
-        float sum = 0.f;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            const uint8_t h = x[i].qh[in+j] >> im;
-            sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16))
-                 + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16))
-                 + y[j+32] * d * s[2] * ((q[j+ 0] >>  4) - ((h >> 4) & 1 ? 0 : 16))
-                 + y[j+48] * d * s[3] * ((q[j+16] >>  4) - ((h >> 6) & 1 ? 0 : 16));
-        }
-        tmp += sum;
-    }
-#endif
-
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
 
@@ -470,8 +352,6 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx,
 
     const block_q6_K * x = (const block_q6_K *)vx + ib0;
 
-#if QK_K == 256
-
     const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
     const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
 
@@ -526,37 +406,6 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx,
 
     }
 
-#else
-
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0...3
-
-    const int step = tid * K_QUANTS_PER_ITERATION;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y  = yy + i * QK_K + step;
-        const uint8_t * ql = x[i].ql + step;
-        const uint8_t * qh = x[i].qh + step;
-        const int8_t  * s  = x[i].scales;
-
-        const float d = x[i+0].d;
-
-        float sum = 0;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32)
-                 + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32)
-                 + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >>  4) | ((qh[j] & 0x30) >> 0)) - 32)
-                 + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >>  4) | ((qh[j] & 0xc0) >> 2)) - 32);
-        }
-        tmp += sum;
-
-    }
-
-#endif
-
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
 

ggml-cuda/mmq.cu

@@ -826,11 +826,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
         const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd;
 
-#if QK_K == 256
         x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm;
-#else
-        x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]};
-#endif
     }
 
 #pragma unroll
@@ -933,9 +929,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
         const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd;
 
-#if QK_K == 256
         x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm;
-#endif
     }
 
 #pragma unroll

ggml-cuda/vecdotq.cuh

@@ -712,7 +712,6 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
 static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 
-#ifndef GGML_QKK_64
     const block_q4_K * bq4_K = (const block_q4_K *) vbq;
 
     int    v[2];
@@ -754,58 +753,11 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
     }
 
     return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8);
-
-#else
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    const block_q4_K * bq4_K = (const block_q4_K *) vbq;
-
-    float sumf_d = 0.0f;
-    float sumf_m = 0.0f;
-
-    uint16_t aux16[2];
-    const uint8_t * s = (const uint8_t *)aux16;
-
-    const uint16_t * a = (const uint16_t *)bq4_K->scales;
-    aux16[0] = a[0] & 0x0f0f;
-    aux16[1] = (a[0] >> 4) & 0x0f0f;
-
-    const float dall = bq4_K->dm[0];
-    const float dmin = bq4_K->dm[1];
-
-    const float d8_1 = __low2float(bq8_1[0].ds);
-    const float d8_2 = __low2float(bq8_1[1].ds);
-
-    const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
-    const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
-    const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
-    const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
-
-    const int * q4 = (const int *)bq4_K->qs + (iqs/2);
-    const int v1 = q4[0];
-    const int v2 = q4[4];
-
-    const int dot1 = __dp4a(ui2, v2 & 0x0f0f0f0f, __dp4a(ui1, v1 & 0x0f0f0f0f, 0));
-    const int dot2 = __dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, __dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0));
-    const int dot3 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0));
-    const int dot4 = __dp4a(0x01010101, ui4, __dp4a(0x01010101, ui3, 0));
-
-    sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]);
-    sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]);
-
-    return dall * sumf_d - dmin * sumf_m;
-
-#else
-    NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-
-#endif
 }
 
 static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 
-#ifndef GGML_QKK_64
     const block_q5_K * bq5_K = (const block_q5_K *) vbq;
 
     int   vl[2];
@@ -847,48 +799,6 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
     }
 
     return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8);
-
-#else
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    const block_q5_K * bq5_K = (const block_q5_K *) vbq;
-
-    const int8_t * s = bq5_K->scales;
-
-    const float d = bq5_K->d;
-
-    const float d8_1 = __low2half(bq8_1[0].ds);
-    const float d8_2 = __low2half(bq8_1[1].ds);
-
-    const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
-    const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
-    const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
-    const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
-
-    const int * ql = (const int *)bq5_K->qs + (iqs/2);
-    const int vl1 = ql[0];
-    const int vl2 = ql[4];
-
-    const int step = 4 * (iqs/2); // 0, 4, 8, 12
-    const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6
-    const int in = step%8; // 0, 4, 0, 4
-    const int vh = (*((const int *)(bq5_K->qh + in))) >> im;
-
-    const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f);
-    const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f);
-    const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f);
-    const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f);
-
-    const float sumf_d = d8_1 * (__dp4a(ui1, v1, 0) * s[0] + __dp4a(ui2, v2, 0) * s[1])
-                       + d8_2 * (__dp4a(ui3, v3, 0) * s[2] + __dp4a(ui4, v4, 0) * s[3]);
-
-    return d * sumf_d;
-
-#else
-    NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-
-#endif
 }
 
 static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
@@ -919,7 +829,6 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
 
 static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-#if QK_K == 256
     const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq;
 
 #if QR2_XXS == 8
@@ -960,15 +869,11 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
     }
     return d * (sumi1 + sumi2);
 #endif
-#else
-    NO_DEVICE_CODE;
-#endif
 }
 
 static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-#if QK_K == 256
     const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq;
 
     const int ib32 = iqs;
@@ -1002,17 +907,12 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
     GGML_UNUSED(ksigns64);
     NO_DEVICE_CODE;
 #endif
-#else
-    GGML_UNUSED(ksigns64);
-    NO_DEVICE_CODE;
-#endif
 }
 
 // TODO
 static __device__ __forceinline__ float vec_dot_iq2_s_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-#if QK_K == 256
     const block_iq2_s * bq2 = (const block_iq2_s *) vbq;
 
     const int ib32 = iqs;
@@ -1048,16 +948,11 @@ static __device__ __forceinline__ float vec_dot_iq2_s_q8_1(
     GGML_UNUSED(ksigns64);
     NO_DEVICE_CODE;
 #endif
-#else
-    GGML_UNUSED(ksigns64);
-    NO_DEVICE_CODE;
-#endif
 }
 
 static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-#if QK_K == 256
     const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq;
 
     const int ib32 = iqs;
@@ -1082,16 +977,12 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
 #else
     NO_DEVICE_CODE;
 #endif
-#else
-    NO_DEVICE_CODE;
-#endif
 }
 
 // TODO: don't use lookup table for signs
 static __device__ __forceinline__ float vec_dot_iq3_s_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-#if QK_K == 256
     const block_iq3_s * bq2 = (const block_iq3_s *) vbq;
 
     const int ib32 = iqs;
@@ -1114,14 +1005,10 @@ static __device__ __forceinline__ float vec_dot_iq3_s_q8_1(
 #else
     NO_DEVICE_CODE;
 #endif
-#else
-    NO_DEVICE_CODE;
-#endif
 }
 
 static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-#if QK_K == 256
     const block_iq1_s * bq1 = (const block_iq1_s *) vbq;
 
     const int ib32 = iqs;
@@ -1149,14 +1036,10 @@ static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
     const float d = d1q * __low2float (bq8_1[ib32].ds);
     const float m = d1q * __high2float(bq8_1[ib32].ds);
     return d * sumi + m * delta;
-#else
-    NO_DEVICE_CODE;
-#endif
 }
 
 static __device__ __forceinline__ float vec_dot_iq1_m_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-#if QK_K == 256
     const block_iq1_m * bq1 = (const block_iq1_m *) vbq;
 
     const int ib32 = iqs;
@@ -1192,9 +1075,6 @@ static __device__ __forceinline__ float vec_dot_iq1_m_q8_1(
     scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
     const float d = (float)scale.f16 * __low2float (bq8_1[ib32].ds);
     return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1));
-#else
-    NO_DEVICE_CODE;
-#endif
 }
 
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
@@ -1250,9 +1130,7 @@ static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1(
 static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 
-#if QK_K == 256
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-
     const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq;
     const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
 
@@ -1270,10 +1148,6 @@ static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
         sumi2 = __dp4a(v2, q8[j+4], sumi2);
     }
     return d * (sumi1 + sumi2);
-
-#else
-    NO_DEVICE_CODE;
-#endif
 #else
     return vec_dot_iq4_xs_q8_1(vbq, bq8_1, iqs);
 #endif

ggml-impl.h

@@ -144,6 +144,10 @@ extern "C" {
 #endif
 #endif
 
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+#endif
+
 // 16-bit float
 // on Arm, we use __fp16
 // on x86, we use uint16_t

ggml-metal.m

@@ -35,6 +35,10 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_ROW,
     GGML_METAL_KERNEL_TYPE_DIV,
     GGML_METAL_KERNEL_TYPE_DIV_ROW,
+    GGML_METAL_KERNEL_TYPE_REPEAT_F32,
+    GGML_METAL_KERNEL_TYPE_REPEAT_F16,
+    GGML_METAL_KERNEL_TYPE_REPEAT_I32,
+    GGML_METAL_KERNEL_TYPE_REPEAT_I16,
     GGML_METAL_KERNEL_TYPE_SCALE,
     GGML_METAL_KERNEL_TYPE_SCALE_4,
     GGML_METAL_KERNEL_TYPE_CLAMP,
@@ -184,9 +188,9 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,
-    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,
+  //GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,     // https://github.com/ggerganov/llama.cpp/issues/7261
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,
-    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256,
+  //GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, // https://github.com/ggerganov/llama.cpp/issues/7261
     GGML_METAL_KERNEL_TYPE_CPY_F32_F16,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
     GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,
@@ -381,10 +385,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
                 // dictionary of preprocessor macros
                 NSMutableDictionary * prep = [NSMutableDictionary dictionary];
 
-#ifdef GGML_QKK_64
-                prep[@"GGML_QKK_64"] = @(1);
-#endif
-
                 MTLCompileOptions* options = [MTLCompileOptions new];
                 options.preprocessorMacros = prep;
 
@@ -489,6 +489,10 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW,                       mul_row,                        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV,                           div,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW,                       div_row,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F32,                    repeat_f32,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F16,                    repeat_f16,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I32,                    repeat_i32,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I16,                    repeat_i16,                     true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE,                         scale,                          true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4,                       scale_4,                        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP,                         clamp,                          true);
@@ -638,9 +642,9 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,        flash_attn_ext_f16_h96,         ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,       flash_attn_ext_f16_h112,        ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,       flash_attn_ext_f16_h128,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,       flash_attn_ext_f16_h256,        ctx->support_simdgroup_mm);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,       flash_attn_ext_f16_h256,        ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,   flash_attn_ext_vec_f16_h128,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256,   flash_attn_ext_vec_f16_h256,    ctx->support_simdgroup_reduction);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256,   flash_attn_ext_vec_f16_h256,    ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16,                   cpy_f32_f16,                    true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,                   cpy_f32_f32,                    true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,                  cpy_f32_q8_0,                   true);
@@ -750,6 +754,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
         case GGML_OP_ACC:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
+        case GGML_OP_REPEAT:
         case GGML_OP_SCALE:
         case GGML_OP_CLAMP:
         case GGML_OP_SQR:
@@ -774,6 +779,9 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
         case GGML_OP_LEAKY_RELU:
             return true;
         case GGML_OP_FLASH_ATTN_EXT:
+            if (op->src[0]->ne[0] == 256) {
+                return false;
+            }
             return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
@@ -980,8 +988,6 @@ static enum ggml_status ggml_metal_graph_compute(
             switch (dst->op) {
                 case GGML_OP_CONCAT:
                     {
-                        const int64_t nb = ne00;
-
                         id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CONCAT].pipeline;
 
                         [encoder setComputePipelineState:pipeline];
@@ -1012,7 +1018,6 @@ static enum ggml_status ggml_metal_graph_compute(
                         [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
                         [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
                         [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
-                        [encoder setBytes:&nb   length:sizeof(nb)   atIndex:27];
 
                         const int nth = MIN(1024, ne0);
 
@@ -1022,11 +1027,14 @@ static enum ggml_status ggml_metal_graph_compute(
                 case GGML_OP_MUL:
                 case GGML_OP_DIV:
                     {
+                        GGML_ASSERT(src0t == GGML_TYPE_F32);
+                        GGML_ASSERT(src1t == GGML_TYPE_F32);
+
                         const size_t offs = 0;
 
                         bool bcast_row = false;
 
-                        int64_t nb = ne00;
+                        int64_t nb = ne00; // used by the "row" kernels
 
                         id<MTLComputePipelineState> pipeline = nil;
 
@@ -1095,6 +1103,42 @@ static enum ggml_status ggml_metal_graph_compute(
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                         }
                     } break;
+                case GGML_OP_REPEAT:
+                    {
+                        id<MTLComputePipelineState> pipeline;
+
+                        switch (src0t) {
+                            case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F32].pipeline; break;
+                            case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F16].pipeline; break;
+                            case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I32].pipeline; break;
+                            case GGML_TYPE_I16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I16].pipeline; break;
+                            default: GGML_ASSERT(false);
+                        }
+
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                        [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                        [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                        [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                        [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                        [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
+                        [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                        [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                        [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                        [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
+                        [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
+                        [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
+                        [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
+                        [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
+                        [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
+                        [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
+                        [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
+                        [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
+
+                        const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
+
+                        [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                    } break;
                 case GGML_OP_ACC:
                     {
                         GGML_ASSERT(src0t == GGML_TYPE_F32);
@@ -1773,11 +1817,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_Q3_K) {
-#ifdef GGML_QKK_64
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-#else
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-#endif
                             }
                             else if (src0t == GGML_TYPE_Q5_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -2018,12 +2058,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                     {
                                         nth0 = 4;
                                         nth1 = 16;
-                                    #if QK_K == 64
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
-                                    #else
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline;
-                                    #endif
-
                                     } break;
                                 default:
                                     {
@@ -2088,11 +2123,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_Q3_K) {
-#ifdef GGML_QKK_64
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-#else
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-#endif
                             }
                             else if (src0t == GGML_TYPE_Q5_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -2590,7 +2621,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 case 96:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break;
                                 case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break;
                                 case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break;
-                                case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break;
+                              //case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break;
                                 default:
                                           {
                                               GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
@@ -2603,7 +2634,7 @@ static enum ggml_status ggml_metal_graph_compute(
 
                             switch (ne00) {
                                 case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128].pipeline; break;
-                                case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break;
+                              //case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break;
                                 default:
                                           {
                                               GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);

ggml-metal.metal

@@ -168,6 +168,53 @@ kernel void kernel_div(
     }
 }
 
+template<typename T>
+kernel void kernel_repeat(
+        device const char * src0,
+        device       char * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant uint64_t & nb00,
+        constant uint64_t & nb01,
+        constant uint64_t & nb02,
+        constant uint64_t & nb03,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant uint64_t & nb0,
+        constant uint64_t & nb1,
+        constant uint64_t & nb2,
+        constant uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3 % ne03;
+    const int64_t i02 = i2 % ne02;
+    const int64_t i01 = i1 % ne01;
+
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    device       char * dst_ptr  = dst  +  i3*nb3  +  i2*nb2  +  i1*nb1 ;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        const int i00 = i0 % ne00;
+        *((device T *)(dst_ptr + i0*nb0)) = *((device T *)(src0_ptr + i00*nb00));
+    }
+}
+
+typedef decltype(kernel_repeat<float>) kernel_repeat_t;
+
+template [[host_name("kernel_repeat_f32")]] kernel kernel_repeat_t kernel_repeat<float>;
+template [[host_name("kernel_repeat_f16")]] kernel kernel_repeat_t kernel_repeat<half>;
+template [[host_name("kernel_repeat_i32")]] kernel kernel_repeat_t kernel_repeat<int>;
+template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat<short>;
+
 // assumption: src1 is a row
 // broadcast src1 into src0
 kernel void kernel_add_row(
@@ -2418,7 +2465,7 @@ template [[host_name("kernel_flash_attn_ext_f16_h80" )]] kernel flash_attn_ext_f
 template [[host_name("kernel_flash_attn_ext_f16_h96" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<96>;
 template [[host_name("kernel_flash_attn_ext_f16_h112")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<112>;
 template [[host_name("kernel_flash_attn_ext_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<128>;
-template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<256>;
+//template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<256>;
 
 template<int64_t D, int64_t Q = 1, int64_t C = 32> // head size, queries per threadgroup, cache items per threadgroup
 kernel void kernel_flash_attn_ext_vec_f16(
@@ -2696,7 +2743,7 @@ kernel void kernel_flash_attn_ext_vec_f16(
 }
 
 template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<128>;
-template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<256>;
+//template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<256>;
 
 kernel void kernel_cpy_f16_f16(
         device  const half * src0,
@@ -3386,7 +3433,6 @@ void kernel_mul_mv_q2_K_f32_impl(
 
     const int step = sizeof(block_q2_K) * nb;
 
-#if QK_K == 256
     const int ix = tiisg/8;  // 0...3
     const int it = tiisg%8;  // 0...7
     const int iq = it/4;     // 0 or 1
@@ -3438,57 +3484,6 @@ void kernel_mul_mv_q2_K_f32_impl(
 
         y4 += 4 * QK_K;
     }
-#else
-    const int ix = tiisg/2;  // 0...15
-    const int it = tiisg%2;  // 0...1
-
-    device const float * y4 = y + ix * QK_K + 8 * it;
-
-    for (int ib = ix; ib < nb; ib += 16) {
-
-        float4 sumy = {0.f, 0.f, 0.f, 0.f};
-        for (int i = 0; i < 8; ++i) {
-            yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
-            yl[i+ 8] = y4[i+16]; sumy[1] += yl[i+ 8];
-            yl[i+16] = y4[i+32]; sumy[2] += yl[i+16];
-            yl[i+24] = y4[i+48]; sumy[3] += yl[i+24];
-        }
-
-        device const uint8_t  * sc = (device const uint8_t  *)x[ib].scales;
-        device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
-        device const half     * dh = &x[ib].d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
-            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
-            for (int i = 0; i < 8; i += 2) {
-                acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
-                acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
-                acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
-                acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
-                acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
-                acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
-                acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
-                acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
-            }
-
-            float dall = dh[0];
-            float dmin = dh[1];
-            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
-                                 (acc1[1] + 1.f/256.f * acc2[1]) * (sc[1] & 0xF) * 1.f/ 4.f +
-                                 (acc1[2] + 1.f/256.f * acc2[2]) * (sc[2] & 0xF) * 1.f/16.f +
-                                 (acc1[3] + 1.f/256.f * acc2[3]) * (sc[3] & 0xF) * 1.f/64.f) -
-                         dmin * (sumy[0] * (sc[0] >> 4) + sumy[1] * (sc[1] >> 4) + sumy[2] * (sc[2] >> 4) + sumy[3] * (sc[3] >> 4));
-
-            qs += step/2;
-            sc += step;
-            dh += step/2;
-        }
-
-        y4 += 16 * QK_K;
-    }
-#endif
 
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
@@ -3526,7 +3521,6 @@ kernel void kernel_mul_mv_q2_K_f32(
     kernel_mul_mv_q2_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
 }
 
-#if QK_K == 256
 void kernel_mul_mv_q3_K_f32_impl(
         device const  void * src0,
         device const float * src1,
@@ -3685,84 +3679,6 @@ void kernel_mul_mv_q3_K_f32_impl(
         }
     }
 }
-#else
-void kernel_mul_mv_q3_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    const int nb = ne00/QK_K;
-
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-    const int64_t im = tgpig.z;
-
-    const int row = 2 * r0 + sgitg;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_q3_K * x = (device const block_q3_K *) src0 + row*nb + offset0;
-    device const float     * yy = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
-
-    const int ix = tiisg/4;
-    const int il = 4 * (tiisg%4);// 0, 4, 8, 12
-    const int iq = il/8;         // 0, 0, 1, 1
-    const int in = il%8;         // 0, 4, 0, 4
-
-    float2 sum = {0.f, 0.f};
-
-    for (int i = ix; i < nb; i += 8) {
-
-        const float d_all = (float)(x[i].d);
-
-        device const uint16_t * q = (device const uint16_t *)(x[i].qs + il);
-        device const uint16_t * h = (device const uint16_t *)(x[i].hmask + in);
-        device const uint16_t * s = (device const uint16_t *)(x[i].scales);
-        device const float    * y = yy + i * QK_K + il;
-
-        const float d1 = d_all * ((int32_t)(s[0] & 0x000F) - 8);
-        const float d2 = d_all * ((int32_t)(s[0] & 0x00F0) - 128) * 1.f/64.f;
-        const float d3 = d_all * ((int32_t)(s[0] & 0x0F00) - 2048) * 1.f/4096.f;
-        const float d4 = d_all * ((int32_t)(s[0] & 0xF000) - 32768) * 1.f/262144.f;
-
-        for (int l = 0; l < 4; l += 2) {
-            const uint16_t hm = h[l/2] >> iq;
-            sum[0] += y[l+ 0] * d1 * ((int32_t)(q[l/2] & 0x0003) - ((hm & 0x0001) ? 0 :  4))
-                    + y[l+16] * d2 * ((int32_t)(q[l/2] & 0x000c) - ((hm & 0x0004) ? 0 : 16))
-                    + y[l+32] * d3 * ((int32_t)(q[l/2] & 0x0030) - ((hm & 0x0010) ? 0 : 64))
-                    + y[l+48] * d4 * ((int32_t)(q[l/2] & 0x00c0) - ((hm & 0x0040) ? 0 : 256));
-            sum[1] += y[l+ 1] * d1 * ((int32_t)(q[l/2] & 0x0300) - ((hm & 0x0100) ? 0 : 1024))
-                    + y[l+17] * d2 * ((int32_t)(q[l/2] & 0x0c00) - ((hm & 0x0400) ? 0 : 4096))
-                    + y[l+33] * d3 * ((int32_t)(q[l/2] & 0x3000) - ((hm & 0x1000) ? 0 : 16384))
-                    + y[l+49] * d4 * ((int32_t)(q[l/2] & 0xc000) - ((hm & 0x4000) ? 0 : 65536));
-        }
-
-    }
-    const float sumf = sum[0] + sum[1] * 1.f/256.f;
-
-    const float tot = simd_sum(sumf);
-    if (tiisg == 0) {
-        dst[r1*ne0 + im*ne0*ne1 + row] = tot;
-    }
-
-}
-#endif
 
 [[host_name("kernel_mul_mv_q3_K_f32")]]
 kernel void kernel_mul_mv_q3_K_f32(
@@ -3792,7 +3708,6 @@ kernel void kernel_mul_mv_q3_K_f32(
     kernel_mul_mv_q3_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
 }
 
-#if QK_K == 256
 void kernel_mul_mv_q4_K_f32_impl(
         device const  void * src0,
         device const float * src1,
@@ -3906,103 +3821,6 @@ void kernel_mul_mv_q4_K_f32_impl(
         }
     }
 }
-#else
-void kernel_mul_mv_q4_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    const int ix = tiisg/4;  // 0...7
-    const int it = tiisg%4;  // 0...3
-
-    const int nb = ne00/QK_K;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-    const int first_row = r0 * N_DST;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0;
-    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[8];
-    float yh[8];
-    float sumf[N_DST]={0.f}, all_sum;
-
-    const int step = sizeof(block_q4_K) * nb / 2;
-
-    device const float * y4 = y + ix * QK_K + 8 * it;
-
-    uint16_t sc16[4];
-
-    for (int ib = ix; ib < nb; ib += 8) {
-
-        float2 sumy = {0.f, 0.f};
-        for (int i = 0; i < 8; ++i) {
-            yl[i] = y4[i+ 0]; sumy[0] += yl[i];
-            yh[i] = y4[i+32]; sumy[1] += yh[i];
-        }
-
-        device const uint16_t * sc = (device const uint16_t *)x[ib].scales;
-        device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
-        device const half     * dh = x[ib].d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            sc16[0] = sc[0] & 0x000f;
-            sc16[1] = sc[0] & 0x0f00;
-            sc16[2] = sc[0] & 0x00f0;
-            sc16[3] = sc[0] & 0xf000;
-
-            float2 acc1 = {0.f, 0.f};
-            float2 acc2 = {0.f, 0.f};
-            for (int i = 0; i < 8; i += 2) {
-                acc1[0] += yl[i+0] * (qs[i/2] & 0x000F);
-                acc1[1] += yl[i+1] * (qs[i/2] & 0x0F00);
-                acc2[0] += yh[i+0] * (qs[i/2] & 0x00F0);
-                acc2[1] += yh[i+1] * (qs[i/2] & 0xF000);
-            }
-
-            float dall = dh[0];
-            float dmin = dh[1];
-            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc16[0] +
-                                 (acc2[0] + 1.f/256.f * acc2[1]) * sc16[1] * 1.f/4096.f) -
-                         dmin * 1.f/16.f * (sumy[0] * sc16[2] + sumy[1] * sc16[3] * 1.f/256.f);
-
-            qs += step;
-            sc += step;
-            dh += step;
-        }
-
-        y4 += 8 * QK_K;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
-        }
-    }
-}
-#endif
 
 [[host_name("kernel_mul_mv_q4_K_f32")]]
 kernel void kernel_mul_mv_q4_K_f32(
@@ -4070,8 +3888,6 @@ void kernel_mul_mv_q5_K_f32_impl(
 
     const int step = sizeof(block_q5_K) * nb;
 
-#if QK_K == 256
-#
     float yl[16], yh[16];
 
     const uint16_t kmask1 = 0x3f3f;
@@ -4154,54 +3970,6 @@ void kernel_mul_mv_q5_K_f32_impl(
         y1 += 4 * QK_K;
 
     }
-#else
-    float yl[8], yh[8];
-
-    const int il = 4 * (tiisg/8);  // 0, 4, 8, 12
-    const int ix = tiisg%8;
-    const int iq = il/8;         // 0, 0, 1, 1
-    const int in = il%8;         // 0, 4, 0, 4
-
-    device const float * y = yy + ix*QK_K + il;
-
-    for (int i = ix; i < nb; i += 8) {
-
-        for (int l = 0; l < 4; ++l) {
-            yl[l+0] = y[l+ 0];
-            yl[l+4] = y[l+16];
-            yh[l+0] = y[l+32];
-            yh[l+4] = y[l+48];
-        }
-
-        device const half * dh = &x[i].d;
-        device const uint8_t * q = x[i].qs + il;
-        device const uint8_t * h = x[i].qh + in;
-        device const int8_t  * s = x[i].scales;
-
-        for (int row = 0; row < 2; ++row) {
-
-            const float d = dh[0];
-
-            float2 acc = {0.f, 0.f};
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t hl = h[l] >> iq;
-                acc[0] += yl[l+0] * s[0] * ((int16_t)(q[l+ 0] & 0x0F) - (hl & 0x01 ? 0 : 16))
-                        + yl[l+4] * s[1] * ((int16_t)(q[l+16] & 0x0F) - (hl & 0x04 ? 0 : 16));
-                acc[1] += yh[l+0] * s[2] * ((int16_t)(q[l+ 0] & 0xF0) - (hl & 0x10 ? 0 : 256))
-                        + yh[l+4] * s[3] * ((int16_t)(q[l+16] & 0xF0) - (hl & 0x40 ? 0 : 256));
-            }
-            sumf[row] += d * (acc[0] + 1.f/16.f * acc[1]);
-
-            q += step;
-            h += step;
-            s += step;
-            dh += step/2;
-
-        }
-
-        y += 8 * QK_K;
-    }
-#endif
 
     for (int row = 0; row < 2; ++row) {
         const float tot = simd_sum(sumf[row]);
@@ -4280,7 +4048,6 @@ void kernel_mul_mv_q6_K_f32_impl(
 
     float sumf = 0;
 
-#if QK_K == 256
     const int tid  = tiisg/2;
     const int ix   = tiisg%2;
     const int ip   = tid/8;         // 0 or 1
@@ -4316,30 +4083,6 @@ void kernel_mul_mv_q6_K_f32_impl(
 
     }
 
-#else
-    const int ix  = tiisg/4;
-    const int il  = 4*(tiisg%4);
-
-    for (int i = ix; i < nb; i += 8) {
-        device const float * y = yy + i * QK_K + il;
-        device const uint8_t * ql = x[i].ql + il;
-        device const uint8_t * qh = x[i].qh + il;
-        device const int8_t  * s  = x[i].scales;
-
-        const float d = x[i].d;
-
-        float4 sums = {0.f, 0.f, 0.f, 0.f};
-        for (int l = 0; l < 4; ++l) {
-            sums[0] += y[l+ 0] * ((int8_t)((ql[l+ 0] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
-            sums[1] += y[l+16] * ((int8_t)((ql[l+16] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
-            sums[2] += y[l+32] * ((int8_t)((ql[l+ 0] >>  4) | ((qh[l] & kmask3) >> 0)) - 32);
-            sums[3] += y[l+48] * ((int8_t)((ql[l+16] >>  4) | ((qh[l] & kmask4) >> 2)) - 32);
-        }
-        sumf += d * (sums[0] * s[0] + sums[1] * s[1] + sums[2] * s[2] + sums[3] * s[3]);
-    }
-
-#endif
-
     const float tot = simd_sum(sumf);
     if (tiisg == 0) {
         dst[r1*ne0 + im*ne0*ne1 + row] = tot;
@@ -5173,9 +4916,7 @@ void kernel_mul_mv_iq1_m_f32_impl(
 
     device const float * y4 = y + 32 * ix;
 
-#if QK_K != 64
     iq1m_scale_t scale;
-#endif
 
     for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
 
@@ -5196,10 +4937,7 @@ void kernel_mul_mv_iq1_m_f32_impl(
         device const uint16_t * sc = (device const uint16_t *)xr->scales;
 
         for (int row = 0; row < N_DST; row++) {
-
-#if QK_K != 64
             scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
-#endif
 
             constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
             constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
@@ -5215,14 +4953,9 @@ void kernel_mul_mv_iq1_m_f32_impl(
             }
             const float delta1 = sumy[0] * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA) + sumy[1] * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
             const float delta2 = sumy[2] * (qh[1] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA) + sumy[3] * (qh[1] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
-#if QK_K == 64
-            const float d = (float) *((device const half *)(sc - 1));
-            sumf[row] += d * ((sum[0] + delta1) * (2*((sc[0] >> (8*(ib%2)+0)) & 0xf) + 1) +
-                              (sum[1] + delta2) * (2*((sc[0] >> (8*(ib%2)+4)) & 0xf) + 1));
-#else
+
             sumf[row] += (float)scale.f16 * ((sum[0] + delta1) * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 7) + 1) +
                                              (sum[1] + delta2) * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 7) + 1));
-#endif
 
             sc += nb*sizeof(block_iq1_m)/2;
             qs += nb*sizeof(block_iq1_m);
@@ -5334,7 +5067,6 @@ void kernel_mul_mv_iq4_nl_f32_impl(
     }
 }
 
-#if QK_K != 64
 void kernel_mul_mv_iq4_xs_f32_impl(
         device const  void * src0,
         device const float * src1,
@@ -5429,7 +5161,6 @@ void kernel_mul_mv_iq4_xs_f32_impl(
         }
     }
 }
-#endif
 
 [[host_name("kernel_mul_mv_iq1_s_f32")]]
 kernel void kernel_mul_mv_iq1_s_f32(
@@ -5542,11 +5273,7 @@ kernel void kernel_mul_mv_iq4_xs_f32(
         uint tiisg[[thread_index_in_simdgroup]],
         uint sgitg[[simdgroup_index_in_threadgroup]]) {
 
-#if QK_K == 64
-    kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
-#else
     kernel_mul_mv_iq4_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
-#endif
 }
 
 //============================= templates and their specializations =============================
@@ -5672,10 +5399,9 @@ void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg
     float dl, ml;
     uint8_t sc = xb->scales[il];
 
-#if QK_K == 256
     q = q + 32*(il/8) + 16*(il&1);
     il = (il/2)%4;
-#endif
+
     half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
     uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
     dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
@@ -5691,7 +5417,6 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg
     device const uint8_t * h = (device const uint8_t *)xb->hmask;
     device const int8_t * scales = (device const int8_t *)xb->scales;
 
-#if QK_K == 256
     q = q + 32 * (il/8) + 16 * (il&1);
     h = h + 16 * (il&1);
     uint8_t m = 1 << (il/2);
@@ -5712,17 +5437,6 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg
     for (int i = 0; i < 16; ++i) {
         reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
     }
-#else
-    float    kcoef = il&1 ? 1.f/16.f : 1.f;
-    uint16_t kmask = il&1 ? 0xF0     : 0x0F;
-    float    dl = d_all * ((scales[il/2] & kmask) * kcoef - 8);
-    float    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
-    uint8_t  mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
-    uint8_t  m = 1<<(il*2);
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = coef * dl * ((q[i] & mask) - ((h[i%8] & (m * (1 + i/8))) ? 0 : 4.f/coef));
-    }
-#endif
 }
 
 static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
@@ -5734,7 +5448,6 @@ template <typename type4x4>
 void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) {
     device const uchar * q = xb->qs;
 
-#if QK_K == 256
     short is = (il/4) * 2;
     q = q + (il/4) * 32 + 16 * (il&1);
     il = il & 3;
@@ -5743,16 +5456,7 @@ void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg
     const float min = xb->dmin;
     const float dl = d * sc[0];
     const float ml = min * sc[1];
-#else
-    (void) get_scale_min_k4_just2;
 
-    q = q + 16 * (il&1);
-    device const uint8_t * s = xb->scales;
-    device const half2 * dh = (device const half2 *)xb->d;
-    const float2 d = (float2)dh[0];
-    const float dl = il<2 ? d[0] * (s[0]&0xF) : d[0] * (s[1]&0xF)/16.h;
-    const float ml = il<2 ? d[1] * (s[0]>>4)  : d[1] * (s[1]>>4);
-#endif
     const ushort mask = il<2 ? 0x0F : 0xF0;
     for (int i = 0; i < 16; ++i) {
         reg[i/4][i%4] = dl * (q[i] & mask) - ml;
@@ -5764,7 +5468,6 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
     device const uint8_t * q  = xb->qs;
     device const uint8_t * qh = xb->qh;
 
-#if QK_K == 256
     short is = (il/4) * 2;
     q  = q + 32 * (il/4) + 16 * (il&1);
     qh = qh + 16 * (il&1);
@@ -5781,17 +5484,6 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
     for (int i = 0; i < 16; ++i) {
         reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
     }
-#else
-    q = q + 16 * (il&1);
-    device const int8_t * s = xb->scales;
-    const float dl = xb->d * s[il];
-    uint8_t m = 1<<(il*2);
-    const float  coef = il<2 ? 1.f  : 1.f/16.f;
-    const ushort mask = il<2 ? 0x0F : 0xF0;
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = coef * dl * ((q[i] & mask) - (qh[i%8] & (m*(1+i/8)) ? 0.f : 16.f/coef));
-    }
-#endif
 }
 
 template <typename type4x4>
@@ -5801,15 +5493,11 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg
     device const uint8_t * qh = (device const uint8_t *)xb->qh;
     device const int8_t * scales = (device const int8_t *)xb->scales;
 
-#if QK_K == 256
     ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
     qh = qh + 32*(il/8) + 16*(il&1);
     float sc = scales[(il%2) + 2 * ((il/2))];
     il = (il/2) & 3;
-#else
-    ql = ql + 16 * (il&1);
-    float sc = scales[il];
-#endif
+
     const uint16_t  kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
     const uint16_t  kmask2 = il>1 ? 0xF0              : 0x0F;
     const float       coef = il>1 ? 1.f/16.f          : 1.f;
@@ -5966,20 +5654,15 @@ void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 &
     const int ib32 = il/2;
     il = il%2;
     device const uint16_t * sc = (device const uint16_t *)xb->scales;
-#if QK_K == 64
-    const float d = xb->d;
-#else
+
     iq1m_scale_t scale;
     scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
     const float d = scale.f16;
-#endif
+
     device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
     device const uint8_t * qh = xb->qh + 2*ib32 + il;
-#if QK_K == 64
-    const float dl  = d * (2*((sc[ib32/2] >> (8*(ib32%2)+4*il)) & 0xf) + 1);
-#else
+
     const float dl  = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
-#endif
     const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
     const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
     constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
@@ -6009,9 +5692,6 @@ void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4
 
 template <typename type4x4>
 void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
-#if QK_K == 64
-    dequantize_iq4_nl(xb, il, reg);
-#else
     // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
     const int ib32 = il/2;
     il = il%2;
@@ -6028,7 +5708,6 @@ void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4
         reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
         reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
     }
-#endif
 }
 
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
@@ -6533,11 +6212,7 @@ kernel void kernel_mul_mm_id(
         sgitg);
 }
 
-#if QK_K == 256
 #define QK_NL 16
-#else
-#define QK_NL 4
-#endif
 
 //
 // get rows
@@ -6577,11 +6252,7 @@ template [[host_name("kernel_get_rows_iq2_s")]]   kernel get_rows_t kernel_get_r
 template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_t kernel_get_rows<block_iq1_s,   QK_NL, dequantize_iq1_s>;
 template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_t kernel_get_rows<block_iq1_m,   QK_NL, dequantize_iq1_m>;
 template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_t kernel_get_rows<block_iq4_nl,  2,     dequantize_iq4_nl>;
-#if QK_K == 64
-template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_t kernel_get_rows<block_iq4_xs,  2,     dequantize_iq4_xs>;
-#else
 template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_t kernel_get_rows<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
-#endif
 
 //
 // matrix-matrix multiplication
@@ -6609,11 +6280,7 @@ template [[host_name("kernel_mul_mm_iq2_s_f32")]]   kernel mat_mm_t kernel_mul_m
 template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq1_s,   QK_NL, dequantize_iq1_s>;
 template [[host_name("kernel_mul_mm_iq1_m_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq1_m,   QK_NL, dequantize_iq1_m>;
 template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_nl,  2,     dequantize_iq4_nl>;
-#if QK_K == 64
-template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_nl,  2,     dequantize_iq4_xs>;
-#else
 template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
-#endif
 
 //
 // indirect matrix-matrix multiplication
@@ -6641,11 +6308,7 @@ template [[host_name("kernel_mul_mm_id_iq2_s_f32")]]   kernel mat_mm_id_t kernel
 template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_s,   QK_NL, dequantize_iq1_s>;
 template [[host_name("kernel_mul_mm_id_iq1_m_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_m,   QK_NL, dequantize_iq1_m>;
 template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_nl,  2,     dequantize_iq4_nl>;
-#if QK_K == 64
-template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs,  2,     dequantize_iq4_xs>;
-#else
 template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
-#endif
 
 //
 // matrix-vector multiplication
@@ -6854,7 +6517,5 @@ template [[host_name("kernel_mul_mv_id_iq3_xxs_f32")]] kernel kernel_mul_mv_id_t
 template [[host_name("kernel_mul_mv_id_iq3_s_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq3_s_f32_impl>>;
 template [[host_name("kernel_mul_mv_id_iq2_s_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_s_f32_impl>>;
 template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_nl_f32_impl>>;
-#if QK_K != 64
 template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_xs_f32_impl>>;
-#endif
 

ggml-opencl.cpp

@@ -1,4 +1,4 @@
-#include "ggml.h"
+#include "ggml.h"
 #include "ggml-opencl.h"
 #include "ggml-backend-impl.h"
 

ggml.c

@@ -871,22 +871,14 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
     },
     [GGML_TYPE_IQ4_XS] = {
         .type_name                = "iq4_xs",
-#if QK_K == 64
-        .blck_size                = QK4_NL,
-#else
         .blck_size                = QK_K,
-#endif
         .type_size                = sizeof(block_iq4_xs),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_iq4_xs,
         .from_float               = quantize_row_iq4_xs,
         .from_float_reference     = (ggml_from_float_t)quantize_row_iq4_xs_reference,
         .vec_dot                  = ggml_vec_dot_iq4_xs_q8_K,
-#if QK_K == 64
-        .vec_dot_type             = GGML_TYPE_Q8_0,
-#else
         .vec_dot_type             = GGML_TYPE_Q8_K,
-#endif
         .nrows                    = 1,
     },
     [GGML_TYPE_Q8_K] = {
@@ -2678,9 +2670,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "ARGSORT",
     "LEAKY_RELU",
 
-    "FLASH_ATTN",
     "FLASH_ATTN_EXT",
-    "FLASH_FF",
     "FLASH_ATTN_BACK",
     "SSM_CONV",
     "SSM_SCAN",
@@ -2706,7 +2696,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CROSS_ENTROPY_LOSS_BACK",
 };
 
-static_assert(GGML_OP_COUNT == 76, "GGML_OP_COUNT != 76");
+static_assert(GGML_OP_COUNT == 74, "GGML_OP_COUNT != 74");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -2768,9 +2758,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "argsort(x)",
     "leaky_relu(x)",
 
-    "flash_attn(x)",
     "flash_attn_ext(x)",
-    "flash_ff(x)",
     "flash_attn_back(x)",
     "ssm_conv(x)",
     "ssm_scan(x)",
@@ -2796,7 +2784,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "cross_entropy_loss_back(x,y)",
 };
 
-static_assert(GGML_OP_COUNT == 76, "GGML_OP_COUNT != 76");
+static_assert(GGML_OP_COUNT == 74, "GGML_OP_COUNT != 74");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -6956,38 +6944,6 @@ struct ggml_tensor * ggml_top_k(
     return result;
 }
 
-// ggml_flash_attn
-
-struct ggml_tensor * ggml_flash_attn(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * q,
-        struct ggml_tensor  * k,
-        struct ggml_tensor  * v,
-        bool                  masked) {
-    GGML_ASSERT(ggml_can_mul_mat(k, q));
-    // TODO: check if vT can be multiplied by (k*qT)
-
-    bool is_node = false;
-
-    if (q->grad || k->grad || v->grad) {
-        is_node = true;
-    }
-
-    //struct ggml_tensor * result = ggml_dup_tensor(ctx, q);
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, q->ne);
-
-    int32_t t = masked ? 1 : 0;
-    ggml_set_op_params(result, &t, sizeof(t));
-
-    result->op   = GGML_OP_FLASH_ATTN;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = q;
-    result->src[1] = k;
-    result->src[2] = v;
-
-    return result;
-}
-
 // ggml_flash_attn_ext
 
 struct ggml_tensor * ggml_flash_attn_ext(
@@ -7047,38 +7003,6 @@ void ggml_flash_attn_ext_set_prec(
     ggml_set_op_params_i32(a, 2, prec_i32); // scale is on first pos, max_bias on second
 }
 
-// ggml_flash_ff
-
-struct ggml_tensor * ggml_flash_ff(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        struct ggml_tensor  * b0,
-        struct ggml_tensor  * b1,
-        struct ggml_tensor  * c0,
-        struct ggml_tensor  * c1) {
-    GGML_ASSERT(ggml_can_mul_mat(b0, a));
-    // TODO: more checks
-
-    bool is_node = false;
-
-    if (a->grad || b0->grad || b1->grad || c0->grad || c1->grad) {
-        is_node = true;
-    }
-
-    //struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, a->ne);
-
-    result->op   = GGML_OP_FLASH_FF;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-    result->src[1] = b0;
-    result->src[2] = b1;
-    result->src[3] = c0;
-    result->src[4] = c1;
-
-    return result;
-}
-
 // ggml_flash_attn_back
 
 struct ggml_tensor * ggml_flash_attn_back(
@@ -7088,6 +7012,8 @@ struct ggml_tensor * ggml_flash_attn_back(
         struct ggml_tensor  * v,
         struct ggml_tensor  * d,
         bool                  masked) {
+    GGML_ASSERT(false && "TODO: adapt to ggml_flash_attn_ext() changes");
+
     GGML_ASSERT(ggml_can_mul_mat(k, q));
     // TODO: check if vT can be multiplied by (k*qT)
 
@@ -15717,400 +15643,6 @@ static void ggml_compute_forward_argsort(
     }
 }
 
-// ggml_compute_forward_flash_attn
-
-static void ggml_compute_forward_flash_attn_f32(
-        const struct ggml_compute_params * params,
-        const bool masked,
-        struct ggml_tensor * dst) {
-
-    const struct ggml_tensor * q = dst->src[0];
-    const struct ggml_tensor * k = dst->src[1];
-    const struct ggml_tensor * v = dst->src[2];
-
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
-
-    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
-    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
-    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
-    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int64_t D = neq0;
-    const int64_t N = neq1;
-    const int64_t P = nek1 - N;
-    const int64_t M = P + N;
-
-    const int Mup = ggml_up(M, GGML_SOFT_MAX_UNROLL);
-
-    GGML_ASSERT(ne0 == D);
-    GGML_ASSERT(ne1 == N);
-    GGML_ASSERT(P >= 0);
-
-    GGML_ASSERT(nbq0 == sizeof(float));
-    GGML_ASSERT(nbk0 == sizeof(float));
-    GGML_ASSERT(nbv0 == sizeof(float));
-
-    GGML_ASSERT(neq0 == D);
-    GGML_ASSERT(nek0 == D);
-    GGML_ASSERT(nev1 == D);
-
-    GGML_ASSERT(neq1 == N);
-    GGML_ASSERT(nek1 == N + P);
-    GGML_ASSERT(nev1 == D);
-
-    // dst cannot be transposed or permuted
-    GGML_ASSERT(nb0 == sizeof(float));
-    GGML_ASSERT(nb0 <= nb1);
-    GGML_ASSERT(nb1 <= nb2);
-    GGML_ASSERT(nb2 <= nb3);
-
-    if (params->type == GGML_TASK_TYPE_INIT) {
-        return;
-    }
-
-    if (params->type == GGML_TASK_TYPE_FINALIZE) {
-        return;
-    }
-
-    // parallelize by q rows using ggml_vec_dot_f32
-
-    // total rows in q
-    const int nr = neq1*neq2*neq3;
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
-    const float scale = 1.0f/sqrtf(D);
-
-    //printf("P=%d N=%d D=%d ir0=%d ir1=%d scale = %f\n", P, N, D, ir0, ir1, scale);
-
-    for (int ir = ir0; ir < ir1; ++ir) {
-        // q indices
-        const int iq3 = ir/(neq2*neq1);
-        const int iq2 = (ir - iq3*neq2*neq1)/neq1;
-        const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1);
-
-        float * S = (float *) params->wdata + ith*(Mup + CACHE_LINE_SIZE_F32);
-
-        for (int i = M; i < Mup; ++i) {
-            S[i] = -INFINITY;
-        }
-
-        const int64_t masked_begin = masked ? (P + iq1 + 1) : M;
-        for (int64_t ic = 0; ic < masked_begin; ++ic) {
-            // k indices
-            const int ik3 = iq3;
-            const int ik2 = iq2 % nek2;
-            const int ik1 = ic;
-
-            // S indices
-            const int i1 = ik1;
-
-            ggml_vec_dot_f32(neq0,
-                    S + i1, 0,
-                    (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)), 0,
-                    (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)), 0, 1);
-        }
-
-        // scale
-        ggml_vec_scale_f32(masked_begin, S, scale);
-
-        for (int64_t i = masked_begin; i < M; i++) {
-            S[i] = -INFINITY;
-        }
-
-        // softmax
-        // exclude known -INF S[..] values from max and loop
-        // dont forget to set their SW values to zero
-        {
-            float max = -INFINITY;
-            ggml_vec_max_f32(masked_begin, &max, S);
-
-            ggml_float sum = 0.0;
-            {
-#ifdef GGML_SOFT_MAX_ACCELERATE
-                max = -max;
-                vDSP_vsadd(S, 1, &max, S, 1, Mup);
-                vvexpf(S, S, &Mup);
-                ggml_vec_sum_f32(Mup, &sum, S);
-#else
-                sum = ggml_vec_soft_max_f32(Mup, S, S, max);
-#endif
-            }
-
-            assert(sum > 0.0);
-
-            sum = 1.0/sum;
-            ggml_vec_scale_f32(masked_begin, S, sum);
-
-#ifndef NDEBUG
-            for (int i = 0; i < masked_begin; ++i) {
-                assert(!isnan(S[i]));
-                assert(!isinf(S[i]));
-            }
-#endif
-        }
-
-        for (int64_t ic = 0; ic < nev1; ++ic) {
-            // dst indices
-            const int i1 = iq1;
-            const int i2 = iq2;
-            const int i3 = iq3;
-
-            // v indices
-            const int iv2 = iq2 % nev2;
-            const int iv3 = iq3;
-
-            ggml_vec_dot_f32(masked_begin,
-                    (float *) ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2   + i3*nb3)), 0,
-                    (float *) ((char *) v->data   + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)), 0,
-                    S, 0, 1);
-        }
-    }
-}
-
-static void ggml_compute_forward_flash_attn_f16(
-        const struct ggml_compute_params * params,
-        const bool masked,
-        struct ggml_tensor * dst) {
-
-    const struct ggml_tensor * q = dst->src[0];
-    const struct ggml_tensor * k = dst->src[1];
-    const struct ggml_tensor * v = dst->src[2];
-
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
-
-    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
-    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
-    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
-    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int64_t D = neq0;
-    const int64_t N = neq1;
-    const int64_t P = nek1 - N;
-    const int64_t M = P + N;
-
-    const int Mup = ggml_up(M, GGML_SOFT_MAX_UNROLL);
-
-    GGML_ASSERT(ne0 == D);
-    GGML_ASSERT(ne1 == N);
-    GGML_ASSERT(P >= 0);
-
-    GGML_ASSERT(nbq0 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nbk0 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nbv0 == sizeof(ggml_fp16_t));
-
-    GGML_ASSERT(neq0 == D);
-    GGML_ASSERT(nek0 == D);
-    GGML_ASSERT(nev1 == D);
-
-    GGML_ASSERT(neq1 == N);
-    GGML_ASSERT(nek1 == N + P);
-    GGML_ASSERT(nev1 == D);
-
-    // dst cannot be transposed or permuted
-    GGML_ASSERT(nb0 == sizeof(float));
-    GGML_ASSERT(nb0 <= nb1);
-    GGML_ASSERT(nb1 <= nb2);
-    GGML_ASSERT(nb2 <= nb3);
-
-    if (params->type == GGML_TASK_TYPE_INIT) {
-        return;
-    }
-
-    if (params->type == GGML_TASK_TYPE_FINALIZE) {
-        return;
-    }
-
-    // parallelize by q rows using ggml_vec_dot_f32
-
-    // total rows in q
-    const int nr = neq1*neq2*neq3;
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
-    const float scale = 1.0f/sqrtf(D);
-
-    //printf("P=%d N=%d D=%d ir0=%d ir1=%d scale = %f\n", P, N, D, ir0, ir1, scale);
-
-    for (int ir = ir0; ir < ir1; ++ir) {
-        // q indices
-        const int iq3 = ir/(neq2*neq1);
-        const int iq2 = (ir - iq3*neq2*neq1)/neq1;
-        const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1);
-
-        float * S = (float *) params->wdata + ith*(2*Mup + CACHE_LINE_SIZE_F32);
-
-        for (int i = M; i < Mup; ++i) {
-            S[i] = -INFINITY;
-        }
-
-        if (GGML_VEC_DOT_UNROLL > 2 || nek1 % GGML_VEC_DOT_UNROLL != 0) {
-            for (int64_t ic = 0; ic < nek1; ++ic) {
-                // k indices
-                const int ik3 = iq3;
-                const int ik2 = iq2 % nek2;
-                const int ik1 = ic;
-
-                // S indices
-                const int i1 = ik1;
-
-                ggml_vec_dot_f16(neq0,
-                        S + i1, 0,
-                        (ggml_fp16_t *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)), 0,
-                        (ggml_fp16_t *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)), 0, 1);
-            }
-        } else {
-            for (int64_t ic = 0; ic < nek1; ic += GGML_VEC_DOT_UNROLL) {
-                // k indices
-                const int ik3 = iq3;
-                const int ik2 = iq2 % nek2;
-                const int ik1 = ic;
-
-                // S indices
-                const int i1 = ik1;
-
-                ggml_vec_dot_f16_unroll(neq0, nbk1,
-                        S + i1,
-                        ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
-                        (ggml_fp16_t *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
-            }
-        }
-
-        // scale
-        ggml_vec_scale_f32(nek1, S, scale);
-
-        if (masked) {
-            for (int64_t i = P; i < M; i++) {
-                if (i > P + iq1) {
-                    S[i] = -INFINITY;
-                }
-            }
-        }
-
-        // softmax
-        // todo: exclude known -INF S[..] values from max and loop, assuming their results to be zero.
-        // dont forget to set their S values to zero
-        {
-            float max = -INFINITY;
-            ggml_vec_max_f32(M, &max, S);
-
-            ggml_float sum = 0.0;
-            {
-#ifdef GGML_SOFT_MAX_ACCELERATE
-                max = -max;
-                vDSP_vsadd(S, 1, &max, S, 1, Mup);
-                vvexpf(S, S, &Mup);
-                ggml_vec_sum_f32(Mup, &sum, S);
-#else
-                sum = ggml_vec_soft_max_f32(Mup, S, S, max);
-#endif
-            }
-
-            assert(sum > 0.0);
-
-            sum = 1.0/sum;
-            ggml_vec_scale_f32(M, S, sum);
-
-#ifndef NDEBUG
-            for (int i = 0; i < M; ++i) {
-                assert(!isnan(S[i]));
-                assert(!isinf(S[i]));
-            }
-#endif
-        }
-
-        ggml_fp16_t * S16 = (ggml_fp16_t *) ((float *) params->wdata + ith*(2*Mup + CACHE_LINE_SIZE_F32) + Mup);
-
-        for (int64_t i = 0; i < M; i++) {
-            S16[i] = GGML_FP32_TO_FP16(S[i]);
-        }
-
-        // todo: exclude known zero S[..] values from dot (reducing nev0 and increasing begin of v and S16).
-        if (GGML_VEC_DOT_UNROLL == 1 || (nev1 % GGML_VEC_DOT_UNROLL != 0)) {
-            for (int64_t ic = 0; ic < nev1; ++ic) {
-                // dst indices
-                const int i1 = iq1;
-                const int i2 = iq2;
-                const int i3 = iq3;
-
-                // v indices
-                const int iv2 = iq2 % nev2;
-                const int iv3 = iq3;
-
-                ggml_vec_dot_f16(nev0,
-                        (float *)       ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2   + i3*nb3)), 0,
-                        (ggml_fp16_t *) ((char *) v->data   + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)), 0,
-                        S16, 0, 1);
-            }
-        } else {
-            for (int64_t ic = 0; ic < nev1; ic += GGML_VEC_DOT_UNROLL) {
-                // dst indices
-                const int i1 = iq1;
-                const int i2 = iq2;
-                const int i3 = iq3;
-
-                // v indices
-                const int iv2 = iq2 % nev2;
-                const int iv3 = iq3;
-
-                ggml_vec_dot_f16_unroll(nev0, nbv1,
-                        (float *) ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2   + i3*nb3)),
-                        ((char *)             v->data + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)),
-                        S16);
-            }
-        }
-    }
-}
-
-static void ggml_compute_forward_flash_attn(
-        const struct ggml_compute_params * params,
-        const bool masked,
-        struct ggml_tensor * dst) {
-
-    const struct ggml_tensor * q = dst->src[0];
-
-    switch (q->type) {
-        case GGML_TYPE_F16:
-            {
-                ggml_compute_forward_flash_attn_f16(params, masked, dst);
-            } break;
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_flash_attn_f32(params, masked, dst);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
-    }
-}
-
 // ggml_compute_forward_flash_attn_ext
 
 static void ggml_compute_forward_flash_attn_ext_f16(
@@ -16344,165 +15876,6 @@ static void ggml_compute_forward_flash_attn_ext(
     }
 }
 
-// ggml_compute_forward_flash_ff
-
-static void ggml_compute_forward_flash_ff_f16(
-        const struct ggml_compute_params * params,
-        struct ggml_tensor * dst) {
-
-    const struct ggml_tensor * a = dst->src[0];  // F16
-    const struct ggml_tensor * b0 = dst->src[1]; // F16 fc_w
-    const struct ggml_tensor * b1 = dst->src[2]; // F32 fc_b
-    const struct ggml_tensor * c0 = dst->src[3]; // F16 proj_w
-    const struct ggml_tensor * c1 = dst->src[4]; // F32 proj_b
-
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
-
-    GGML_TENSOR_LOCALS(int64_t, nea,  a,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nba,  a,   nb)
-    GGML_TENSOR_LOCALS(int64_t, neb0, b0,  ne)
-    GGML_TENSOR_LOCALS(size_t,  nbb0, b0,  nb)
-    GGML_TENSOR_LOCALS(int64_t, neb1, b1,  ne)
-    GGML_TENSOR_LOCALS(size_t,  nbb1, b1,  nb)
-    GGML_TENSOR_LOCALS(int64_t, nec0, c0,  ne)
-    GGML_TENSOR_LOCALS(size_t,  nbc0, c0,  nb)
-    GGML_TENSOR_LOCALS(int64_t, nec1, c1,  ne)
-    GGML_TENSOR_LOCALS(size_t,  nbc1, c1,  nb)
-    GGML_TENSOR_LOCALS(int64_t, ne,   dst, ne)
-    GGML_TENSOR_LOCALS(size_t,  nb,   dst, nb)
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int64_t D = nea0;
-    //const int64_t N = nea1;
-    const int64_t M = neb01;
-
-    GGML_ASSERT(ne0 == nea0);
-    GGML_ASSERT(ne1 == nea1);
-    GGML_ASSERT(ne2 == nea2);
-
-    GGML_ASSERT(nba0  == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nbb00 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nbb10 == sizeof(float));
-    GGML_ASSERT(nbc00 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nbc10 == sizeof(float));
-
-    GGML_ASSERT(neb00 == D);
-    GGML_ASSERT(neb01 == M);
-    GGML_ASSERT(neb10 == M);
-    GGML_ASSERT(neb11 == 1);
-
-    GGML_ASSERT(nec00 == M);
-    GGML_ASSERT(nec01 == D);
-    GGML_ASSERT(nec10 == D);
-    GGML_ASSERT(nec11 == 1);
-
-    // dst cannot be transposed or permuted
-    GGML_ASSERT(nb0 == sizeof(float));
-    GGML_ASSERT(nb0 <= nb1);
-    GGML_ASSERT(nb1 <= nb2);
-    GGML_ASSERT(nb2 <= nb3);
-
-    if (params->type == GGML_TASK_TYPE_INIT) {
-        return;
-    }
-
-    if (params->type == GGML_TASK_TYPE_FINALIZE) {
-        return;
-    }
-
-    // parallelize by a rows using ggml_vec_dot_f32
-
-    // total rows in a
-    const int nr = nea1*nea2*nea3;
-
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
-
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
-    for (int ir = ir0; ir < ir1; ++ir) {
-        // a indices
-        const int ia3 = ir/(nea2*nea1);
-        const int ia2 = (ir - ia3*nea2*nea1)/nea1;
-        const int ia1 = (ir - ia3*nea2*nea1 - ia2*nea1);
-
-        float * S = (float *) params->wdata + ith*(2*M + CACHE_LINE_SIZE_F32);
-
-        for (int64_t ic = 0; ic < neb01; ++ic) {
-            // b0 indices
-            const int ib03 = ia3;
-            const int ib02 = ia2;
-            const int ib01 = ic;
-
-            // S indices
-            const int i1 = ib01;
-
-            ggml_vec_dot_f16(nea0,
-                    S + i1, 0,
-                    (ggml_fp16_t *) ((char *) b0->data + (ib01*nbb01 + ib02*nbb02 + ib03*nbb03)), 0,
-                    (ggml_fp16_t *) ((char *)  a->data + ( ia1*nba1  +  ia2*nba2  +  ia3*nba3)), 0, 1);
-        }
-
-        ggml_vec_add_f32(neb01, S, S, (float *) b1->data);
-        //ggml_vec_gelu_f32(neb01, S, S);
-
-        ggml_fp16_t * S16 = (ggml_fp16_t *) ((float *) params->wdata + ith*(2*M + CACHE_LINE_SIZE_F32) + M);
-
-        for (int64_t i = 0; i < M; i++) {
-            S16[i] = GGML_FP32_TO_FP16(S[i]);
-        }
-
-        ggml_vec_gelu_f16(neb01, S16, S16);
-
-        {
-            // dst indices
-            const int i1 = ia1;
-            const int i2 = ia2;
-            const int i3 = ia3;
-
-            for (int64_t ic = 0; ic < nec01; ++ic) {
-
-                ggml_vec_dot_f16(neb01,
-                        (float *)       ((char *) dst->data + (ic*nb0 + i1*nb1   + i2*nb2   + i3*nb3)), 0,
-                        (ggml_fp16_t *) ((char *) c0->data  + (         ic*nbc01 + i2*nbc02 + i3*nbc03)), 0,
-                        S16, 0, 1);
-            }
-
-            ggml_vec_add_f32(nec01,
-                    (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3)),
-                    (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3)),
-                    (float *) c1->data);
-        }
-    }
-}
-
-static void ggml_compute_forward_flash_ff(
-        const struct ggml_compute_params * params,
-        struct ggml_tensor * dst) {
-
-    const struct ggml_tensor * b0 = dst->src[1];
-
-    switch (b0->type) {
-        case GGML_TYPE_F16:
-            {
-                ggml_compute_forward_flash_ff_f16(params, dst);
-            } break;
-        case GGML_TYPE_F32:
-            {
-                GGML_ASSERT(false); // TODO
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
-    }
-}
-
 // ggml_compute_forward_flash_attn_back
 
 static void ggml_compute_forward_flash_attn_back_f32(
@@ -18073,21 +17446,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_leaky_relu(params, tensor);
             } break;
-        case GGML_OP_FLASH_ATTN:
-            {
-                const int32_t t = ggml_get_op_params_i32(tensor, 0);
-                GGML_ASSERT(t == 0 || t == 1);
-                const bool masked = t != 0;
-                ggml_compute_forward_flash_attn(params, masked, tensor);
-            } break;
         case GGML_OP_FLASH_ATTN_EXT:
             {
                 ggml_compute_forward_flash_attn_ext(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3], tensor);
             } break;
-        case GGML_OP_FLASH_FF:
-            {
-                ggml_compute_forward_flash_ff(params, tensor);
-            } break;
         case GGML_OP_FLASH_ATTN_BACK:
             {
                 int32_t t = ggml_get_op_params_i32(tensor, 0);
@@ -19094,7 +18456,6 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             {
                 GGML_ASSERT(false); // TODO: not implemented
             } break;
-        case GGML_OP_FLASH_ATTN:
         case GGML_OP_FLASH_ATTN_EXT:
             {
                 struct ggml_tensor * flash_grad = NULL;
@@ -19148,10 +18509,6 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                             zero_table);
                 }
             } break;
-        case GGML_OP_FLASH_FF:
-            {
-                GGML_ASSERT(false); // not supported
-            } break;
         case GGML_OP_FLASH_ATTN_BACK:
             {
                 GGML_ASSERT(false); // not supported
@@ -19838,15 +19195,10 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads, int n_cur_
             {
                 n_tasks = n_threads;
             } break;
-        case GGML_OP_FLASH_ATTN:
         case GGML_OP_FLASH_ATTN_EXT:
             {
                 n_tasks = n_threads;
             } break;
-        case GGML_OP_FLASH_FF:
-            {
-                n_tasks = n_threads;
-            } break;
         case GGML_OP_FLASH_ATTN_BACK:
             {
                 n_tasks = n_threads;
@@ -20243,40 +19595,12 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
                     cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03;
                     cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12;
                 } break;
-            case GGML_OP_FLASH_ATTN:
-                {
-                    const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL);
-
-                    if (node->src[1]->type == GGML_TYPE_F32) {
-                        cur  = sizeof(float)*ne11*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*ne11*n_tasks; // this is overestimated by x2
-                    } else if (node->src[1]->type == GGML_TYPE_F16) {
-                        cur  = sizeof(float)*ne11*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*ne11*n_tasks; // this is overestimated by x2
-                    } else if (node->src[1]->type == GGML_TYPE_BF16) {
-                        cur  = sizeof(float)*ne11*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*ne11*n_tasks; // this is overestimated by x2
-                    }
-                } break;
             case GGML_OP_FLASH_ATTN_EXT:
                 {
                     const int64_t ne00 = node->src[0]->ne[0]; // D
 
                     cur = 3*sizeof(float)*ne00*n_tasks; // 3x head size/thread
                 } break;
-            case GGML_OP_FLASH_FF:
-                {
-                    if (node->src[1]->type == GGML_TYPE_F32) {
-                        cur  = sizeof(float)*node->src[1]->ne[1]*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*node->src[1]->ne[1]*n_tasks; // this is overestimated by x2
-                    } else if (node->src[1]->type == GGML_TYPE_F16) {
-                        cur  = sizeof(float)*node->src[1]->ne[1]*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*node->src[1]->ne[1]*n_tasks; // this is overestimated by x2
-                    } else if (node->src[1]->type == GGML_TYPE_BF16) {
-                        cur  = sizeof(float)*node->src[1]->ne[1]*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*node->src[1]->ne[1]*n_tasks; // this is overestimated by x2
-                    }
-                } break;
             case GGML_OP_FLASH_ATTN_BACK:
                 {
                     const int64_t    D = node->src[0]->ne[0];
@@ -22117,11 +21441,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_IQ1_S:   result = quantize_iq1_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ1_M:   result = quantize_iq1_m  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_NL:  result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-#if QK_K == 64
-        case GGML_TYPE_IQ4_XS:  result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-#else
         case GGML_TYPE_IQ4_XS:  result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-#endif
         case GGML_TYPE_F16:
             {
                 size_t elemsize = sizeof(ggml_fp16_t);
@@ -23422,6 +22742,16 @@ int ggml_cpu_has_neon(void) {
 #endif
 }
 
+int ggml_cpu_has_sve(void) {
+#if defined(__ARM_FEATURE_SVE)
+    // TODO: Currently, SVE 256 bit is only supported.
+    GGML_ASSERT(svcntb() == QK8_0);
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_arm_fma(void) {
 #if defined(__ARM_FEATURE_FMA)
     return 1;

ggml.h

@@ -481,9 +481,7 @@ extern "C" {
         GGML_OP_ARGSORT,
         GGML_OP_LEAKY_RELU,
 
-        GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_ATTN_EXT,
-        GGML_OP_FLASH_FF,
         GGML_OP_FLASH_ATTN_BACK,
         GGML_OP_SSM_CONV,
         GGML_OP_SSM_SCAN,
@@ -1761,13 +1759,6 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   k);
 
-    GGML_API struct ggml_tensor * ggml_flash_attn(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * q,
-            struct ggml_tensor  * k,
-            struct ggml_tensor  * v,
-            bool                  masked);
-
 #define GGML_KQ_MASK_PAD 32
 
     // q:    [n_embd, n_batch,     n_head,    1]
@@ -1788,6 +1779,7 @@ extern "C" {
             struct ggml_tensor * a,
             enum ggml_prec       prec);
 
+    // TODO: needs to be adapted to ggml_flash_attn_ext
     GGML_API struct ggml_tensor * ggml_flash_attn_back(
            struct ggml_context * ctx,
            struct ggml_tensor  * q,
@@ -1796,14 +1788,6 @@ extern "C" {
            struct ggml_tensor  * d,
            bool                  masked);
 
-    GGML_API struct ggml_tensor * ggml_flash_ff(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b0,
-            struct ggml_tensor  * b1,
-            struct ggml_tensor  * c0,
-            struct ggml_tensor  * c1);
-
     GGML_API struct ggml_tensor * ggml_ssm_conv(
             struct ggml_context * ctx,
             struct ggml_tensor  * s,
@@ -2420,6 +2404,7 @@ extern "C" {
     GGML_API int ggml_cpu_has_avx512_bf16(void);
     GGML_API int ggml_cpu_has_fma        (void);
     GGML_API int ggml_cpu_has_neon       (void);
+    GGML_API int ggml_cpu_has_sve        (void);
     GGML_API int ggml_cpu_has_arm_fma    (void);
     GGML_API int ggml_cpu_has_metal      (void);
     GGML_API int ggml_cpu_has_f16c       (void);

gguf-py/gguf/constants.py

@@ -139,6 +139,7 @@ class MODEL_ARCH(IntEnum):
     COMMAND_R  = auto()
     DBRX       = auto()
     OLMO       = auto()
+    ARCTIC     = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -167,6 +168,7 @@ class MODEL_TENSOR(IntEnum):
     FFN_DOWN           = auto()
     FFN_UP             = auto()
     FFN_ACT            = auto()
+    FFN_NORM_EXP       = auto()
     FFN_GATE_EXP       = auto()
     FFN_DOWN_EXP       = auto()
     FFN_UP_EXP         = auto()
@@ -218,6 +220,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.COMMAND_R:      "command-r",
     MODEL_ARCH.DBRX:           "dbrx",
     MODEL_ARCH.OLMO:           "olmo",
+    MODEL_ARCH.ARCTIC:         "arctic",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -251,6 +254,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.FFN_DOWN_SHEXP:     "blk.{bid}.ffn_down_shexp",
     MODEL_TENSOR.FFN_UP_SHEXP:       "blk.{bid}.ffn_up_shexp",
     MODEL_TENSOR.FFN_ACT:            "blk.{bid}.ffn",
+    MODEL_TENSOR.FFN_NORM_EXP:       "blk.{bid}.ffn_norm_exps",
     MODEL_TENSOR.FFN_GATE_EXP:       "blk.{bid}.ffn_gate_exps",
     MODEL_TENSOR.FFN_DOWN_EXP:       "blk.{bid}.ffn_down_exps",
     MODEL_TENSOR.FFN_UP_EXP:         "blk.{bid}.ffn_up_exps",
@@ -732,6 +736,27 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.ARCTIC: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_ROT_EMBD,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_NORM_EXP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+    ],
     # TODO
 }
 
@@ -905,9 +930,8 @@ class GGUFValueType(IntEnum):
             raise ValueError(f"Unknown type: {type(val)}")
 
 
-# Note: Does not support GGML_QKK_64
-QK_K = 256
 # Items here are (block size, type size)
+QK_K = 256
 GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {
     GGMLQuantizationType.F32:     (1, 4),
     GGMLQuantizationType.F16:     (1, 2),

gguf-py/gguf/gguf_reader.py

@@ -12,6 +12,8 @@ from typing import Any, Literal, NamedTuple, TypeVar, Union
 import numpy as np
 import numpy.typing as npt
 
+from .quants import quant_shape_to_byte_shape
+
 if __name__ == "__main__":
     import sys
     from pathlib import Path
@@ -251,6 +253,7 @@ class GGUFReader:
             tensor_names.add(tensor_name)
             ggml_type = GGMLQuantizationType(raw_dtype[0])
             n_elems = int(np.prod(dims))
+            np_dims = tuple(reversed(dims.tolist()))
             block_size, type_size = GGML_QUANT_SIZES[ggml_type]
             n_bytes = n_elems * type_size // block_size
             data_offs = int(start_offs + offset_tensor[0])
@@ -279,6 +282,7 @@ class GGUFReader:
             else:
                 item_count = n_bytes
                 item_type = np.uint8
+                np_dims = quant_shape_to_byte_shape(np_dims, ggml_type)
             tensors.append(ReaderTensor(
                 name = tensor_name,
                 tensor_type = ggml_type,
@@ -286,7 +290,7 @@ class GGUFReader:
                 n_elements = n_elems,
                 n_bytes = n_bytes,
                 data_offset = data_offs,
-                data = self._get(data_offs, item_type, item_count),
+                data = self._get(data_offs, item_type, item_count).reshape(np_dims),
                 field = field,
             ))
         self.tensors = tensors

gguf-py/gguf/gguf_writer.py

@@ -13,7 +13,6 @@ from string import ascii_letters, digits
 import numpy as np
 
 from .constants import (
-    GGML_QUANT_SIZES,
     GGUF_DEFAULT_ALIGNMENT,
     GGUF_MAGIC,
     GGUF_VERSION,
@@ -26,6 +25,8 @@ from .constants import (
     TokenType,
 )
 
+from .quants import quant_shape_from_byte_shape
+
 logger = logging.getLogger(__name__)
 
 
@@ -229,10 +230,7 @@ class GGUFWriter:
         else:
             dtype = raw_dtype
             if tensor_dtype == np.uint8:
-                block_size, type_size = GGML_QUANT_SIZES[raw_dtype]
-                if tensor_shape[-1] % type_size != 0:
-                    raise ValueError(f"Quantized tensor row size ({tensor_shape[-1]}) is not a multiple of {dtype.name} type size ({type_size})")
-                tensor_shape = tuple(tensor_shape[:-1]) + (tensor_shape[-1] // type_size * block_size,)
+                tensor_shape = quant_shape_from_byte_shape(tensor_shape, raw_dtype)
         n_dims = len(tensor_shape)
         self.ti_data += self._pack("I", n_dims)
         for i in range(n_dims):

gguf-py/gguf/quants.py

@@ -1,5 +1,5 @@
 from __future__ import annotations
-from typing import Callable
+from typing import Callable, Sequence
 
 from numpy.typing import DTypeLike
 
@@ -9,6 +9,20 @@ from .lazy import LazyNumpyTensor
 import numpy as np
 
 
+def quant_shape_to_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType):
+    block_size, type_size = GGML_QUANT_SIZES[quant_type]
+    if shape[-1] % block_size != 0:
+        raise ValueError(f"Quantized tensor row size ({shape[-1]}) is not a multiple of {quant_type.name} block size ({block_size})")
+    return (*shape[:-1], shape[-1] // block_size * type_size)
+
+
+def quant_shape_from_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType):
+    block_size, type_size = GGML_QUANT_SIZES[quant_type]
+    if shape[-1] % type_size != 0:
+        raise ValueError(f"Quantized tensor bytes per row ({shape[-1]}) is not a multiple of {quant_type.name} type size ({type_size})")
+    return (*shape[:-1], shape[-1] // type_size * block_size)
+
+
 # same as ggml_compute_fp32_to_bf16 in ggml-impl.h
 def __compute_fp32_to_bf16(n: np.ndarray) -> np.ndarray:
     n = n.astype(np.float32, copy=False).view(np.int32)

gguf-py/gguf/tensor_mapping.py

@@ -244,6 +244,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
             "model.layers.{bid}.mlp.c_fc",                            # starcoder2
             "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2
+            "model.layers.{bid}.residual_mlp.w3",                     # arctic
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -272,6 +273,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.fc12",              # nomic-bert
             "encoder.layer.{bid}.mlp.gated_layers_w",     # jina-bert-v2
             "transformer.h.{bid}.mlp.linear_1",           # refact
+            "model.layers.{bid}.residual_mlp.w1",         # arctic
         ),
 
         MODEL_TENSOR.FFN_GATE_EXP: (
@@ -306,6 +308,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.fc2",                           # nomic-bert
             "model.layers.{bid}.mlp.c_proj",                          # starcoder2
             "encoder.layer.{bid}.mlp.wo",                             # jina-bert-v2
+            "model.layers.{bid}.residual_mlp.w2",                     # arctic
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -382,6 +385,18 @@ class TensorNameMap:
         ),
     }
 
+    # architecture-specific block mappings
+    arch_block_mappings_cfg: dict[MODEL_ARCH, dict[MODEL_TENSOR, tuple[str, ...]]] = {
+        MODEL_ARCH.ARCTIC: {
+            MODEL_TENSOR.FFN_NORM: (
+                "model.layers.{bid}.residual_layernorm",
+            ),
+            MODEL_TENSOR.FFN_NORM_EXP: (
+                "model.layers.{bid}.post_attention_layernorm",
+            ),
+        },
+    }
+
     mapping: dict[str, tuple[MODEL_TENSOR, str]]
 
     def __init__(self, arch: MODEL_ARCH, n_blocks: int):
@@ -393,12 +408,14 @@ class TensorNameMap:
             self.mapping[tensor_name] = (tensor, tensor_name)
             for key in keys:
                 self.mapping[key] = (tensor, tensor_name)
+        if arch in self.arch_block_mappings_cfg:
+            self.block_mappings_cfg.update(self.arch_block_mappings_cfg[arch])
         for bid in range(n_blocks):
             for tensor, keys in self.block_mappings_cfg.items():
                 if tensor not in MODEL_TENSORS[arch]:
                     continue
                 # TODO: make this configurable
-                n_experts = 60
+                n_experts = 128
                 for xid in range(n_experts):
                     tensor_name = TENSOR_NAMES[tensor].format(bid = bid, xid = xid)
                     self.mapping[tensor_name] = (tensor, tensor_name)

gguf-py/scripts/gguf-new-metadata.py

@@ -118,9 +118,7 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new
 
     for tensor in reader.tensors:
         total_bytes += tensor.n_bytes
-        # Dimensions are written in reverse order, so flip them first
-        shape = np.flipud(tensor.shape).tolist()
-        writer.add_tensor_info(tensor.name, shape, tensor.data.dtype, tensor.data.nbytes, tensor.tensor_type)
+        writer.add_tensor_info(tensor.name, tensor.data.shape, tensor.data.dtype, tensor.data.nbytes, tensor.tensor_type)
 
     bar = tqdm(desc="Writing", total=total_bytes, unit="byte", unit_scale=True)
 

llama.cpp

@@ -26,13 +26,9 @@
 #ifdef GGML_USE_METAL
 #  include "ggml-metal.h"
 #endif
-#ifndef QK_K
-#  ifdef GGML_QKK_64
-#    define QK_K 64
-#  else
-#    define QK_K 256
-#  endif
-#endif
+
+// TODO: replace with ggml API call
+#define QK_K 256
 
 #ifdef __has_include
     #if __has_include(<unistd.h>)
@@ -107,7 +103,7 @@
 #endif
 
 #define LLAMA_MAX_NODES   8192
-#define LLAMA_MAX_EXPERTS 60
+#define LLAMA_MAX_EXPERTS 128
 
 //
 // logging
@@ -225,6 +221,7 @@ enum llm_arch {
     LLM_ARCH_COMMAND_R,
     LLM_ARCH_DBRX,
     LLM_ARCH_OLMO,
+    LLM_ARCH_ARCTIC,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -261,6 +258,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_COMMAND_R,       "command-r"    },
     { LLM_ARCH_DBRX,            "dbrx"         },
     { LLM_ARCH_OLMO,            "olmo"         },
+    { LLM_ARCH_ARCTIC,          "arctic"       },
     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
 };
 
@@ -459,6 +457,7 @@ enum llm_tensor {
     LLM_TENSOR_FFN_DOWN_EXP,  // split experts for backward compatibility
     LLM_TENSOR_FFN_GATE_EXP,
     LLM_TENSOR_FFN_UP_EXP,
+    LLM_TENSOR_FFN_NORM_EXPS,
     LLM_TENSOR_FFN_DOWN_EXPS, // merged experts
     LLM_TENSOR_FFN_GATE_EXPS,
     LLM_TENSOR_FFN_UP_EXPS,
@@ -1036,6 +1035,28 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_ARCTIC,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_NORM_EXPS,   "blk.%d.ffn_norm_exps" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -1696,17 +1717,24 @@ static llama_state g_state;
 // available llama models
 enum e_model {
     MODEL_UNKNOWN,
+    MODEL_14M,
     MODEL_17M,
     MODEL_22M,
     MODEL_33M,
+    MODEL_70M,
     MODEL_109M,
     MODEL_137M,
+    MODEL_160M,
     MODEL_335M,
+    MODEL_410M,
     MODEL_0_5B,
     MODEL_1B,
+    MODEL_1_4B,
     MODEL_2B,
+    MODEL_2_8B,
     MODEL_3B,
     MODEL_4B,
+    MODEL_6_9B,
     MODEL_7B,
     MODEL_8B,
     MODEL_12B,
@@ -1729,6 +1757,7 @@ enum e_model {
     MODEL_8x7B,
     MODEL_8x22B,
     MODEL_16x12B,
+    MODEL_10B_128x3_66B,
 };
 
 static const size_t kiB = 1024;
@@ -1738,6 +1767,7 @@ static const size_t GiB = 1024*MiB;
 struct llama_hparams {
     bool vocab_only;
     bool rope_finetuned;
+    bool use_par_res;
 
     uint32_t n_vocab;
     uint32_t n_ctx_train; // context size the model was trained on
@@ -1903,6 +1933,7 @@ struct llama_layer {
     struct ggml_tensor * ffn_norm_b;
     struct ggml_tensor * layer_out_norm;
     struct ggml_tensor * layer_out_norm_b;
+    struct ggml_tensor * ffn_norm_exps;
 
     // ff
     struct ggml_tensor * ffn_gate; // w1
@@ -2479,7 +2510,6 @@ static bool llama_kv_cache_init(
 static bool llama_kv_cache_find_slot(
            struct llama_kv_cache & cache,
         const struct llama_batch & batch) {
-    const uint32_t n_ctx    = cache.size;
     const uint32_t n_tokens = batch.n_tokens;
 
     if (cache.recurrent) {
@@ -2530,16 +2560,16 @@ static bool llama_kv_cache_find_slot(
     }
     // otherwise, one cell per token.
 
-    if (n_tokens > n_ctx) {
-        LLAMA_LOG_ERROR("%s: n_tokens=%d > n_ctx=%d\n", __func__, n_tokens, n_ctx);
+    if (n_tokens > cache.size) {
+        LLAMA_LOG_ERROR("%s: n_tokens=%d > cache.size=%d\n", __func__, n_tokens, cache.size);
         return false;
     }
 
     uint32_t n_tested = 0;
 
     while (true) {
-        if (cache.head + n_tokens > n_ctx) {
-            n_tested += n_ctx - cache.head;
+        if (cache.head + n_tokens > cache.size) {
+            n_tested += cache.size - cache.head;
             cache.head = 0;
             continue;
         }
@@ -2558,7 +2588,7 @@ static bool llama_kv_cache_find_slot(
             break;
         }
 
-        if (n_tested >= n_ctx) {
+        if (n_tested >= cache.size) {
             //LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens);
             return false;
         }
@@ -3778,40 +3808,48 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
 
 static const char * llama_model_type_name(e_model type) {
     switch (type) {
-        case MODEL_17M:    return "17M";
-        case MODEL_22M:    return "22M";
-        case MODEL_33M:    return "33M";
-        case MODEL_109M:   return "109M";
-        case MODEL_137M:   return "137M";
-        case MODEL_335M:   return "335M";
-        case MODEL_0_5B:   return "0.5B";
-        case MODEL_1B:     return "1B";
-        case MODEL_2B:     return "2B";
-        case MODEL_3B:     return "3B";
-        case MODEL_4B:     return "4B";
-        case MODEL_7B:     return "7B";
-        case MODEL_8B:     return "8B";
-        case MODEL_12B:    return "12B";
-        case MODEL_13B:    return "13B";
-        case MODEL_14B:    return "14B";
-        case MODEL_15B:    return "15B";
-        case MODEL_20B:    return "20B";
-        case MODEL_30B:    return "30B";
-        case MODEL_34B:    return "34B";
-        case MODEL_35B:    return "35B";
-        case MODEL_40B:    return "40B";
-        case MODEL_65B:    return "65B";
-        case MODEL_70B:    return "70B";
-        case MODEL_314B:   return "314B";
-        case MODEL_SMALL:  return "0.1B";
-        case MODEL_MEDIUM: return "0.4B";
-        case MODEL_LARGE:  return "0.8B";
-        case MODEL_XL:     return "1.5B";
-        case MODEL_A2_7B:  return "A2.7B";
-        case MODEL_8x7B:   return "8x7B";
-        case MODEL_8x22B:  return "8x22B";
-        case MODEL_16x12B: return "16x12B";
-        default:           return "?B";
+        case MODEL_14M:           return "14M";
+        case MODEL_17M:           return "17M";
+        case MODEL_22M:           return "22M";
+        case MODEL_33M:           return "33M";
+        case MODEL_70M:           return "70M";
+        case MODEL_109M:          return "109M";
+        case MODEL_137M:          return "137M";
+        case MODEL_160M:          return "160M";
+        case MODEL_335M:          return "335M";
+        case MODEL_410M:          return "410M";
+        case MODEL_0_5B:          return "0.5B";
+        case MODEL_1B:            return "1B";
+        case MODEL_1_4B:          return "1.4B";
+        case MODEL_2B:            return "2B";
+        case MODEL_2_8B:          return "2.8B";
+        case MODEL_3B:            return "3B";
+        case MODEL_4B:            return "4B";
+        case MODEL_6_9B:          return "6.9B";
+        case MODEL_7B:            return "7B";
+        case MODEL_8B:            return "8B";
+        case MODEL_12B:           return "12B";
+        case MODEL_13B:           return "13B";
+        case MODEL_14B:           return "14B";
+        case MODEL_15B:           return "15B";
+        case MODEL_20B:           return "20B";
+        case MODEL_30B:           return "30B";
+        case MODEL_34B:           return "34B";
+        case MODEL_35B:           return "35B";
+        case MODEL_40B:           return "40B";
+        case MODEL_65B:           return "65B";
+        case MODEL_70B:           return "70B";
+        case MODEL_314B:          return "314B";
+        case MODEL_SMALL:         return "0.1B";
+        case MODEL_MEDIUM:        return "0.4B";
+        case MODEL_LARGE:         return "0.8B";
+        case MODEL_XL:            return "1.5B";
+        case MODEL_A2_7B:         return "A2.7B";
+        case MODEL_8x7B:          return "8x7B";
+        case MODEL_8x22B:         return "8x22B";
+        case MODEL_16x12B:        return "16x12B";
+        case MODEL_10B_128x3_66B: return "10B+128x3.66B";
+        default:                  return "?B";
     }
 }
 
@@ -4287,6 +4325,65 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_GPTNEOX:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+                ml.get_key(LLM_KV_USE_PARALLEL_RESIDUAL, hparams.use_par_res);
+                switch (hparams.n_layer) {
+                    case 6:
+                        switch (hparams.n_ff) {
+                            case 512: model.type = e_model::MODEL_14M; break;
+                            case 2048: model.type = e_model::MODEL_70M; break;
+                            default: model.type = e_model::MODEL_UNKNOWN;
+                        } break;
+                    case 12:
+                        switch (hparams.n_ff) {
+                            case 3072: model.type = e_model::MODEL_160M; break;
+                            default: model.type = e_model::MODEL_UNKNOWN;
+                        } break;
+                    case 16:
+                        switch (hparams.n_ff) {
+                            case 8192: model.type = e_model::MODEL_1B; break;
+                            default: model.type = e_model::MODEL_UNKNOWN;
+                        } break;
+                    case 24:
+                        switch (hparams.n_ff) {
+                            case 4096: model.type = e_model::MODEL_410M; break;
+                            case 8192: model.type = e_model::MODEL_1_4B; break;
+                            default: model.type = e_model::MODEL_UNKNOWN;
+                        } break;
+                    case 32:
+                        switch (hparams.n_ff) {
+                            case 10240: model.type = e_model::MODEL_2_8B; break;
+                            case 16384: model.type = e_model::MODEL_6_9B; break;
+                            default: model.type = e_model::MODEL_UNKNOWN;
+                        } break;
+                    case 36:
+                        switch (hparams.n_ff) {
+                            case 20480: model.type = e_model::MODEL_12B; break;
+                            default: model.type = e_model::MODEL_UNKNOWN;
+                        } break;
+                    case 44:
+                        switch (hparams.n_ff) {
+                            case 24576: model.type = e_model::MODEL_20B; break;
+                            default: model.type = e_model::MODEL_UNKNOWN;
+                        } break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_ARCTIC:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                if (hparams.n_expert == 128) {
+                    switch (hparams.n_layer) {
+                        case 35: model.type = e_model::MODEL_10B_128x3_66B; break;
+                        default: model.type = e_model::MODEL_UNKNOWN;
+                    }
+                } else {
+                    model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         default: (void)0;
     }
 
@@ -4496,6 +4593,9 @@ static void llm_load_vocab(
             } else if (
                 tokenizer_pre == "dbrx") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DBRX;
+            } else if (
+                tokenizer_pre == "smaug-bpe") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_SMAUG;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -6038,6 +6138,81 @@ static bool llm_load_tensors(
                         layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
                     }
                 } break;
+            case LLM_ARCH_GPTNEOX:
+                {
+                    model.tok_embd   = ml.create_tensor(ctx_input,  tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab});
+                    // output
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output_norm_b = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+                        layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa});
+
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+                        layer.bo   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd});
+
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+                        layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd});
+
+                        layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd});
+
+                        layer.ffn_up     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff});
+                        layer.ffn_up_b   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff});
+                    }
+                } break;
+            case LLM_ARCH_ARCTIC:
+                {
+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_embd});
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_embd, n_embd});
+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert});
+                        layer.ffn_norm_exps = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM_EXPS, "weight", i), {n_embd});
+                        layer.ffn_gate_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, false);
+                        layer.ffn_down_exps = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert});
+                        layer.ffn_up_exps   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert});
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -6627,7 +6802,7 @@ static struct ggml_tensor * llm_build_kqv(
 
         cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias);
 
-        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3) {
+        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX) {
             ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
         }
 
@@ -6636,7 +6811,7 @@ static struct ggml_tensor * llm_build_kqv(
         struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
         cb(kq, "kq", il);
 
-        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3) {
+        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX) {
             // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
             // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
             ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
@@ -10565,6 +10740,274 @@ struct llm_build_context {
 
         return gf;
     }
+
+    struct ggml_cgraph * build_gptneox() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    model.layers[il].attn_norm_b,
+                    LLM_NORM, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                cb(cur, "bqkv", il);
+
+                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
+            // ffn
+            if (hparams.use_par_res) {
+                // attention and ffn are computed in parallel
+                // x = x + attn(ln1(x)) + ffn(ln2(x))
+
+                struct ggml_tensor * attn_out = cur;
+
+                cur = llm_build_norm(ctx0, inpL, hparams,
+                        model.layers[il].ffn_norm,
+                        model.layers[il].ffn_norm_b,
+                        LLM_NORM, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,
+                        NULL,                      NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+
+                cur = ggml_add(ctx0, cur, inpL);
+                cb(cur, "ffn_out", il);
+
+                inpL = ggml_add(ctx0, cur, attn_out);
+                cb(inpL, "l_out", il);
+            } else {
+                // attention and ffn are computed sequentially
+                // x = x + attn(ln1(x))
+                // x = x + ffn(ln2(x))
+
+                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+                cb(ffn_inp, "ffn_inp", il);
+
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm,
+                        model.layers[il].ffn_norm_b,
+                        LLM_NORM, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,
+                        NULL,                      NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+
+                inpL = ggml_add(ctx0, cur, ffn_inp);
+                cb(inpL, "l_out", il);
+            }
+        }
+
+        cur = llm_build_norm(ctx0, inpL, hparams,
+                model.output_norm,
+                model.output_norm_b,
+                LLM_NORM, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_arctic() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = llm_build_ffn(ctx0, cur,
+                    model.layers[il].ffn_up,   NULL,
+                    model.layers[il].ffn_gate, NULL,
+                    model.layers[il].ffn_down, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            struct ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp);
+            cb(ffn_out, "ffn_out", il);
+
+            // MoE
+            cur = llm_build_norm(ctx0, inpSA, hparams,
+                    model.layers[il].ffn_norm_exps, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm_exps", il);
+
+            cur = llm_build_moe_ffn(ctx0, cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    cb, il);
+            cb(cur, "ffn_moe_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_out);
+            cb(cur, "ffn_out", il);
+
+            ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
+            if (layer_dir != nullptr) {
+                cur = ggml_add(ctx0, cur, layer_dir);
+            }
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
 };
 
 static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
@@ -10775,6 +11218,14 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_olmo();
             } break;
+        case LLM_ARCH_GPTNEOX:
+            {
+                result = llm.build_gptneox();
+            } break;
+        case LLM_ARCH_ARCTIC:
+            {
+                result = llm.build_arctic();
+            } break;
         default:
             GGML_ASSERT(false);
     }
@@ -12064,6 +12515,7 @@ struct llm_tokenizer_bpe {
                         });
                         break;
                     case LLAMA_VOCAB_PRE_TYPE_DBRX:
+                    case LLAMA_VOCAB_PRE_TYPE_SMAUG:
                         word_collection = unicode_regex_split(text, {
                             // same as llama3
                             "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
@@ -14308,8 +14760,6 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
         else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
                 use_more_bits(qs.i_attention_wv, qs.n_attention_wv)) new_type = GGML_TYPE_Q6_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && qs.i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
-        else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) &&
-                (qs.i_attention_wv < qs.n_attention_wv/8 || qs.i_attention_wv >= 7*qs.n_attention_wv/8)) new_type = GGML_TYPE_Q6_K;
         if (qs.model.type == MODEL_70B) {
             // In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
             // 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
@@ -15769,7 +16219,6 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         // these models do not use RoPE
         case LLM_ARCH_GPT2:
         case LLM_ARCH_GPTJ:
-        case LLM_ARCH_GPTNEOX:
         case LLM_ARCH_MPT:
         case LLM_ARCH_REFACT:
         case LLM_ARCH_BLOOM:
@@ -15789,6 +16238,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_XVERSE:
         case LLM_ARCH_COMMAND_R:
         case LLM_ARCH_OLMO:
+        case LLM_ARCH_ARCTIC:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2
@@ -15805,6 +16255,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_PHI3:
         case LLM_ARCH_GEMMA:
         case LLM_ARCH_STARCODER2:
+        case LLM_ARCH_GPTNEOX:
             return LLAMA_ROPE_TYPE_NEOX;
 
         // all model arches should be listed explicitly here
@@ -17183,6 +17634,14 @@ void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_
     ctx->cparams.n_threads_batch = n_threads_batch;
 }
 
+uint32_t llama_n_threads(struct llama_context * ctx) {
+    return ctx->cparams.n_threads;
+}
+
+uint32_t llama_n_threads_batch(struct llama_context * ctx) {
+    return ctx->cparams.n_threads_batch;
+}
+
 void llama_set_abort_callback(struct llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) {
     ctx->abort_callback      = abort_callback;
     ctx->abort_callback_data = abort_callback_data;
@@ -17406,6 +17865,10 @@ bool llama_token_is_eog(const struct llama_model * model, llama_token token) {
     );
 }
 
+bool llama_token_is_control(const struct llama_model * model, llama_token token) {
+    return llama_is_control_token(model->vocab, token);
+}
+
 llama_token llama_token_bos(const struct llama_model * model) {
     return model->vocab.special_bos_id;
 }
@@ -17617,6 +18080,15 @@ static int32_t llama_chat_apply_template_internal(
             }
         }
         // llama2 templates seem to not care about "add_generation_prompt"
+    } else if (tmpl == "phi3" || (tmpl.find("<|assistant|>") != std::string::npos && tmpl.find("<|end|>") != std::string::npos)) {
+        // Phi 3
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>\n" << message->content << "<|end|>\n";
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
     } else if (tmpl == "zephyr" || tmpl.find("<|user|>") != std::string::npos) {
         // zephyr template
         for (auto message : chat) {
@@ -17749,15 +18221,6 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|start_header_id|>assistant<|end_header_id|>\n\n";
         }
-    } else if (tmpl == "phi3" || (tmpl.find("<|assistant|>") != std::string::npos && tmpl.find("<|end|>") != std::string::npos )) {
-        // Phi 3
-        for (auto message : chat) {
-            std::string role(message->role);
-            ss << "<|" << role << "|>\n" << trim(message->content) << "<|end|>\n";
-        }
-        if (add_ass) {
-            ss << "<|assistant|>\n";
-        }
     } else {
         // template not supported
         return -1;
@@ -17882,6 +18345,7 @@ const char * llama_print_system_info(void) {
     s += "AVX512_BF16 = " + std::to_string(ggml_cpu_has_avx512_bf16()) + " | ";
     s += "FMA = "         + std::to_string(ggml_cpu_has_fma())         + " | ";
     s += "NEON = "        + std::to_string(ggml_cpu_has_neon())        + " | ";
+    s += "SVE = "         + std::to_string(ggml_cpu_has_sve())         + " | ";
     s += "ARM_FMA = "     + std::to_string(ggml_cpu_has_arm_fma())     + " | ";
     s += "F16C = "        + std::to_string(ggml_cpu_has_f16c())        + " | ";
     s += "FP16_VA = "     + std::to_string(ggml_cpu_has_fp16_va())     + " | ";

llama.h

@@ -85,6 +85,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_QWEN2          = 11,
         LLAMA_VOCAB_PRE_TYPE_OLMO           = 12,
         LLAMA_VOCAB_PRE_TYPE_DBRX           = 13,
+        LLAMA_VOCAB_PRE_TYPE_SMAUG          = 14,
     };
 
     // note: these values should be synchronized with ggml_rope
@@ -264,6 +265,8 @@ extern "C" {
         bool check_tensors; // validate model tensor data
     };
 
+    // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
+    //       https://github.com/ggerganov/llama.cpp/pull/7544
     struct llama_context_params {
         uint32_t seed;              // RNG seed, -1 for random
         uint32_t n_ctx;             // text context, 0 = from model
@@ -290,14 +293,14 @@ extern "C" {
         ggml_backend_sched_eval_callback cb_eval;
         void * cb_eval_user_data;
 
-        enum ggml_type type_k; // data type for K cache
-        enum ggml_type type_v; // data type for V cache
+        enum ggml_type type_k; // data type for K cache [EXPERIMENTAL]
+        enum ggml_type type_v; // data type for V cache [EXPERIMENTAL]
 
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool logits_all;  // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
         bool embeddings;  // if true, extract embeddings (together with logits)
         bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
-        bool flash_attn;  // whether to use flash attention
+        bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]
 
         // Abort callback
         // if it returns true, execution of llama_decode() will be aborted
@@ -759,6 +762,12 @@ extern "C" {
     // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens)
     LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch);
 
+    // Get the number of threads used for generation of a single token.
+    LLAMA_API uint32_t llama_n_threads(struct llama_context * ctx);
+
+    // Get the number of threads used for prompt and batch processing (multiple token).
+    LLAMA_API uint32_t llama_n_threads_batch(struct llama_context * ctx);
+
     // Set whether to use causal attention or not
     // If set to true, the model will only attend to the past tokens
     LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn);
@@ -817,6 +826,9 @@ extern "C" {
     // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)
     LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token);
 
+    // Identify if Token Id is a control token or a render-able token
+    LLAMA_API bool llama_token_is_control(const struct llama_model * model, llama_token token);
+
     // Special tokens
     LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
     LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence

tests/test-chat-template.cpp

@@ -49,8 +49,14 @@ int main(void) {
         "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif false == true %}{% set loop_messages = messages %}{% set system_message = 'You are Command-R, a brilliant, sophisticated, AI-assistant trained to assist human users by providing thorough responses. You are trained by Cohere.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% if system_message != false %}{{ '<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>' + system_message + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<|START_OF_TURN_TOKEN|><|USER_TOKEN|>' + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% elif message['role'] == 'assistant' %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>'  + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>' }}{% endif %}",
         // Llama-3
         "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}",
-        // Phi-3
-        "{{ bos_token }}{% for message in messages %}{{'<|' + message['role'] + '|>' + ' ' + message['content'] + '<|end|> ' }}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|> ' }}{% else %}{{ eos_token }}{% endif %}"
+        //Phi-3-mini
+        "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') %}{{'<|user|>' + '\n' + message['content'] + '<|end|>' + '\n' + '<|assistant|>' + '\n'}}{% elif (message['role'] == 'assistant') %}{{message['content'] + '<|end|>' + '\n'}}{% endif %}{% endfor %}",
+        //Phi-3-small
+        "{{ bos_token }}{% for message in messages %}{{'<|' + message['role'] + '|>' + '\n' + message['content'] + '<|end|>\n' }}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|>\n' }}{% else %}{{ eos_token }}{% endif %}",
+        //Phi-3-medium
+        "{% for message in messages %}{% if (message['role'] == 'user') %}{{'<|user|>' + '\n' + message['content'] + '<|end|>' + '\n' + '<|assistant|>' + '\n'}}{% elif (message['role'] == 'assistant') %}{{message['content'] + '<|end|>' + '\n'}}{% endif %}{% endfor %}",
+        //Phi-3-vision
+        "{% for message in messages %}{{'<|' + message['role'] + '|>' + '\n' + message['content'] + '<|end|>\n' }}{% endfor %}{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}{{- '<|assistant|>\n' -}}{% endif %}"
     };
     std::vector<std::string> expected_output = {
         // teknium/OpenHermes-2.5-Mistral-7B
@@ -79,8 +85,14 @@ int main(void) {
         "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>You are a helpful assistant<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Hello<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>Hi there<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Who are you<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>I am an assistant<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Another question<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>",
         // Llama 3
         "<|start_header_id|>system<|end_header_id|>\n\nYou are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHello<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nHi there<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWho are you<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nI am an assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nAnother question<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n",
-        // Phi 3
-        "<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\nI am an assistant<|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
+        //Phi-3-mini
+        "<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\n   I am an assistant   <|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
+        //Phi-3-small
+        "<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\n   I am an assistant   <|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
+        //Phi-3-medium
+        "<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\n   I am an assistant   <|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
+        //Phi-3-vision
+        "<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\n   I am an assistant   <|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
     };
     std::vector<char> formatted_chat(1024);
     int32_t res;

tests/test-grad0.cpp

@@ -1515,90 +1515,50 @@ int main(int argc, const char ** argv) {
         }
 
         // flash_attn f32
-        {
-            srand(seed);
-            const int nargs = 3;
+        // TODO: adapt to ggml_flash_attn_ext() changes
+        //{
+        //    srand(seed);
+        //    const int nargs = 3;
 
-            int64_t ne2[4];
+        //    int64_t ne2[4];
 
-            get_random_dims(ne2, 4);
-            int64_t D = ne2[0];
-            int64_t N = ne2[1];
-            int64_t M = ne2[2] + N;
-            int64_t B = ne2[3];
+        //    get_random_dims(ne2, 4);
+        //    int64_t D = ne2[0];
+        //    int64_t N = ne2[1];
+        //    int64_t M = ne2[2] + N;
+        //    int64_t B = ne2[3];
 
-            for (int masked = 0; masked <= 1; ++masked) {
-                for (int ndims = 2; ndims <= 4; ++ndims) {
-                    int max_nrep = (ndims >= 3) ? 2 : 1;
-                    for (int nrep = 1; nrep < max_nrep; ++nrep) {
-                        int64_t neq[4] = { D, N, B*nrep, ne[3] };
-                        int64_t nek[4] = { D, M, B, ne[3] };
-                        int64_t nev[4] = { M, D, B, ne[3] };
-                        if (ndims == 2) {
-                            neq[2] = 1; neq[3] = 1;
-                            nek[2] = 1; nek[3] = 1;
-                            nev[2] = 1; nev[3] = 1;
-                        } else if (ndims == 3) {
-                            neq[3] = 1;
-                            nek[3] = 1;
-                            nev[3] = 1;
-                        }
-                        x[0] = get_random_tensor_f32(ctx0, ndims, neq, -0.1250f, 0.1250f);
-                        x[1] = get_random_tensor_f32(ctx0, ndims, nek, -0.1250f, 0.1250f);
-                        x[2] = get_random_tensor_f32(ctx0, ndims, nev, -0.1250f, 0.1250f);
-                        ggml_set_param(ctx0, x[0]);
-                        ggml_set_param(ctx0, x[1]);
-                        ggml_set_param(ctx0, x[2]);
+        //    for (int masked = 0; masked <= 1; ++masked) {
+        //        for (int ndims = 2; ndims <= 4; ++ndims) {
+        //            int max_nrep = (ndims >= 3) ? 2 : 1;
+        //            for (int nrep = 1; nrep < max_nrep; ++nrep) {
+        //                int64_t neq[4] = { D, N, B*nrep, ne[3] };
+        //                int64_t nek[4] = { D, M, B, ne[3] };
+        //                int64_t nev[4] = { M, D, B, ne[3] };
+        //                if (ndims == 2) {
+        //                    neq[2] = 1; neq[3] = 1;
+        //                    nek[2] = 1; nek[3] = 1;
+        //                    nev[2] = 1; nev[3] = 1;
+        //                } else if (ndims == 3) {
+        //                    neq[3] = 1;
+        //                    nek[3] = 1;
+        //                    nev[3] = 1;
+        //                }
+        //                x[0] = get_random_tensor_f32(ctx0, ndims, neq, -0.1250f, 0.1250f);
+        //                x[1] = get_random_tensor_f32(ctx0, ndims, nek, -0.1250f, 0.1250f);
+        //                x[2] = get_random_tensor_f32(ctx0, ndims, nev, -0.1250f, 0.1250f);
+        //                ggml_set_param(ctx0, x[0]);
+        //                ggml_set_param(ctx0, x[1]);
+        //                ggml_set_param(ctx0, x[2]);
 
-                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
+        //                struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
 
-                        check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
-                    }
-                }
-            }
-        }
+        //                check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
+        //            }
+        //        }
+        //    }
+        //}
 
-        // flash_attn f16, not yet fully implemented
-        if(0)
-        {
-            srand(seed);
-            const int nargs = 3;
-
-            int64_t ne2[4];
-
-            get_random_dims(ne2, 4);
-            int64_t D = ne2[0];
-            int64_t N = ne2[1];
-            int64_t M = ne2[2] + N;
-            int64_t B = ne2[3];
-
-            for (int masked = 0; masked <= 1; ++masked) {
-                for (int ndims = 2; ndims <= 4; ++ndims) {
-                    int64_t neq[4] = { D, N, B, ne[3] };
-                    int64_t nek[4] = { D, M, B, ne[3] };
-                    int64_t nev[4] = { M, D, B, ne[3] };
-                    if (ndims == 2) {
-                        neq[2] = 1; neq[3] = 1;
-                        nek[2] = 1; nek[3] = 1;
-                        nev[2] = 1; nev[3] = 1;
-                    } else if (ndims == 3) {
-                        neq[3] = 1;
-                        nek[3] = 1;
-                        nev[3] = 1;
-                    }
-                    x[0] = get_random_tensor_f16(ctx0, ndims, neq, -0.1250f, 0.1250f);
-                    x[1] = get_random_tensor_f16(ctx0, ndims, nek, -0.1250f, 0.1250f);
-                    x[2] = get_random_tensor_f16(ctx0, ndims, nev, -0.1250f, 0.1250f);
-                    ggml_set_param(ctx0, x[0]);
-                    ggml_set_param(ctx0, x[1]);
-                    ggml_set_param(ctx0, x[2]);
-
-                    struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
-
-                    check_gradient("flash_attn f16", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
-                }
-            }
-        }
         ggml_free(ctx0);
     }